创造者之忧：伊利亚·苏茨克沃的人工智能之旅

这个人帮助创造了现代人工智能。他是 AlexNet 背后的研究人员之一，这项突破永远改变了 AI。他帮助开创了后来成为 GPT 基础的想法。他也是 OpenAI 和 ChatGPT 背后的关键科学家之一。多年来，他的整个职业生涯都专注于一个目标：让机器更聪明。但后来情况发生了变化。随着 AI 变得更强大，Ilya Sutskever 开始问一个不同的问题。AI 和 AGI 的问题是什么？整个问题在于力量。当力量真的很大时，会发生什么？大多数人将 ChatGPT 视为技术奇迹。Ilya 看到了别的东西：一个比任何人预期都更快逼近的未来，一个机器最终可能变得比人类更聪明的未来。而在 2023 年，他做了一件震惊整个科技行业的事情。他反对 Sam Altman，OpenAI 的 CEO，这家他花了多年时间帮助建立的公司。如果我强烈到想要解雇 Sam，那么我认为世界应该知道原因。直到今天，除了少数人之外，没有人确切知道发生了什么。但有一件事是确定的：人工智能越接近成功，它的某些创造者就越担忧。当我们创造出比我们自己更聪明的数字生命时，还会出现一个长期的生存威胁。我们不知道是否能保持控制。这就是那个帮助构建现代 AI 的人的故事，以及他为什么开始害怕接下来可能发生的事情。

This man helped create modern artificial intelligence. He was one of the researchers behind AlexNet, the breakthrough that changed AI forever. He helped pioneer the ideas that would later become the foundations of GPT. And he was one of the key scientists behind OpenAI and ChatGPT. For years, his entire career was focused on one objective: making machines smarter. But then something changed. As AI became more powerful, Ilya Sutskever started asking a different question. What is the problem of AI and AGI? The whole problem is the power. When the power is really big, what's going to happen? Most people saw ChatGPT as a technological miracle. Ilya saw something else: a future that was approaching much faster than anyone expected, a future where machines might eventually become smarter than humans themselves. And in 2023, he did something that shocked the entire technology industry. He turned against Sam Altman, the CEO of OpenAI, the company he had spent years helping build. And if I felt strongly enough to want to, you know, fire Sam, well, I think the world should know what was that reason. To this day, nobody outside a small group of people know exactly what happened. But one thing is certain: the closer artificial intelligence got to working, the more concerned some of its creators became. There is also a longer-term existential threat that will arise when we create digital beings that are more intelligent than ourselves. We have no idea whether we can stay in control. This is the story of the man who helped build modern AI and why he became afraid of what might come next.

在成为人工智能领域最重要的头脑之一之前，Ilya Sutskever 的早期生活跨越了完全不同的世界。他于 1985 年出生在苏联的一个犹太家庭，但他在那里待的时间不长。苏联已经开始解体，仅仅几年后，他的家人离开了这个国家，移居以色列寻求新生活。所以 Ilya 在 1990 年代的耶路撒冷长大。从很小的时候起，他就表现出抽象思维方面的非凡天赋，尤其是数学。但最重要的是，他痴迷于一个问题：智能究竟是如何运作的？当我十几岁的时候，90 年代末的早期青少年时期，我得到的印象是科学根本不知道智能是如何运作的。这个想法让他着迷。大脑如何识别模式？记忆如何工作？神经元如何相互通信？在大多数青少年思考正常生活的时候，Ilya 却痴迷于智能本身的机制。但他不会在耶路撒冷找到这些问题的答案。还在上高中时，他的家人再次搬家，这次是搬到加拿大多伦多。几乎立刻，Ilya 就不是一个普通学生了。根据多个说法，他在加拿大只上了大约一个月的高中，就被直接录取到多伦多大学，就读计算机科学专业。正是在那里，他遇到了将彻底改变他人生轨迹的人——Geoffrey Hinton 教授。

Before becoming one of the most important minds in artificial intelligence, Ilya Sutskever spent his early life moving across completely different worlds. He was born in the Soviet Union in 1985 into a Jewish family, but his stay there didn't last long. The Soviet Union was already beginning to collapse, and just a few years later, his family left the country, moving to Israel in search of a new life. So Ilya grew up in Jerusalem during the 1990s. From an early age, he showed an unusual talent for abstract thinking, especially mathematics. But more than anything else, he became obsessed with one question: How does intelligence actually work? And when I was a teenager, an early teenager in the late '90s, the sense that I got is that science simply did not know how intelligence worked. That idea fascinated him. How does the brain recognize patterns? How does memory work? How do neurons communicate with each other? At a time when most teenagers were thinking about normal life, Ilya was becoming obsessed with the mechanics of intelligence itself. But it wasn't in Jerusalem where he would find the answer to those questions. While still in high school, his family moved once again, this time to Toronto, Canada. And almost immediately it became clear that Ilya wasn't a normal student. According to multiple accounts, he attended high school in Canada for only about a month before being admitted directly into University of Toronto where he enrolled in computer science. And it was there that he would meet the person who would completely change the course of his life, the professor Geoffrey Hinton.

如今，Hinton 被认为是人工智能历史上最有影响力的人物之一。他常被称为 AI 教父。他是 2018 年图灵奖的获得者，该奖项常被称为计算机科学的诺贝尔奖，后来又在 2024 年获得了诺贝尔物理学奖。但当 Ilya 第一次遇到他时，情况看起来截然不同。Geoffrey Hinton 当时只是计算机科学系的一名教授。多年来，Hinton 一直在研究神经网络，这是一种受人类大脑启发的计算系统。但到了 21 世纪初，AI 界的许多人已经开始远离它们。当时许多研究人员认为神经网络已经走到了死胡同。它们被认为太慢、计算成本太高，无法成为人工智能的未来。但 Ilya 决心了解更多。

Today, Hinton is considered one of the most influential figures in the history of artificial intelligence. He's often referred to as the godfather of AI. He is the winner of the Turing Award in 2018, an award often referred to as the Nobel Prize of Computer Science, and later the Nobel Prize in Physics in 2024. But when Ilya first encountered him, this story looked very different. Geoffrey Hinton at the time was just a professor at the faculty of computer science. For years, Hinton had been working on neural networks, computational systems inspired by the human brain. But by the early 2000s, much of the AI world had already started moving away from them. Many researchers at the time believed neural networks had reached a dead end. They were considered too slow and too computationally expensive to ever become the future of artificial intelligence. But Ilya was determined to know more about it.

我大概是在一个周日在办公室里，有人敲门。不是普通的敲门，而是一种急切的敲门声。于是我走过去开门，门口站着这个年轻学生，他说他暑假在炸薯条，但他更愿意在我的实验室工作。于是我说：“那你为什么不预约一下，我们谈谈？”然后 Ilya 说：“现在怎么样？”这大概就是 Ilya 的性格。

I was in my office probably on a Sunday and there was a knock on the door. Not just any knock, but it was a sort of an urgent knock. So I went and answered the door and this was this young student there and he said he was cooking fries over the summer, but he'd rather be working in my lab. And so I said, "Well, why don't you make an appointment and we'll talk?" And so Ilya said, "How about now?" And that sort of was Ilya's character.

于是 Ilya 开始在 Hinton 的实验室工作，开发和研究神经网络。但 Hinton 和 Ilya 究竟在研究什么？当时，大部分 AI 研究都专注于通过将知识显式编码到系统中来构建系统：通过规则、指令和手工制作的世界表征来设计数字智能。但 Hinton 团队对完全不同的东西感兴趣。他们试图回答另一个问题：机器如何自己学习？与其他 AI 研究人员相反，Hinton 和 Ilya 不是试图编程知识，而是试图编程学习，从概念上让智能自行成长。为了理解这个想法，让我们看一个简单的例子。想象一下，你试图教计算机什么是猫。当时，大多数 AI 系统通过手动编写规则来处理这个问题。研究人员可能会告诉机器：“猫有尖耳朵。猫有胡须。猫有特定的脸型。”换句话说，人类必须显式定义机器应该寻找的特征，塑造它们的知识。所以，它们的智能。这在现实变得非常混乱之前是有效的，因为如果猫被部分遮挡了怎么办？如果光线暗怎么办？如果它侧着脸怎么办？突然间，你不再需要几十条规则了。你需要成千上万条，甚至数百万条。而 Hinton 相信有一种完全不同的方法。与其告诉机器猫长什么样，不如直接给它看数百万张猫的图片，让它自己学习模式？这就是神经网络背后的核心思想：不是教机器思考什么，而是教它们如何学习。

So Ilya started working in Hinton's lab, developing and studying neural networks. But what exactly were Hinton and Ilya working on? At the time, much of AI research was focused on building systems by explicitly encoding knowledge into them: designing digital intelligence through rules, instructions, and hand-crafted representations of the world. But the Hinton team was interested in something very different. They were trying to answer another question: How can a machine learn by itself? Opposite to the other AI researchers, Hinton and Ilya weren't trying to program knowledge. They were trying to program learning instead, and conceptually to make the intelligence grow by itself. To understand this idea, let's look at a simple example. Imagine you are trying to teach a computer what a cat is. At the time, most AI systems approached this problem by manually writing rules. A researcher might tell the machine, "Cats have pointy ears. Cats have whiskers. Cats have a certain face shape." In other words, humans had to explicitly define the characteristics that machines should look for, shaping their knowledge. So, their intelligence. And that worked until reality became really messy because what happens if the cat is partially hidden? What if it's dark? What if it's looking sideways? Suddenly, you don't need dozens of rules anymore. You need thousands, maybe millions. And Hinton believed there was a completely different approach. Instead of telling a machine what a cat looks like, what if you simply showed it millions of cat pictures and allowed it to learn patterns on its own? And that was the core idea behind neural networks: not teaching machines what to think, but teaching them how to learn.

我很早就进入了 AI 领域。我对数学和大脑以及“我是什么”非常感兴趣。我是一台计算机。这怎么可能？这很奇怪。所以 AI 非常有趣。其中，机器学习，特别是学习，才是真正神秘的东西，因为在数学中你有逻辑、演绎，但你怎么知道明天太阳会升起？那有什么依据？

I got into AI pretty early. I was very interested in math and brains and what am I? And I'm a computer. How can that be? That's very strange. So AI was very interesting. And within that, machine learning, learning specifically was the really mysterious thing because in math you've got logic, the deduction, but how can you know that the sun's going to rise tomorrow? What's the basis for that?

在 Hinton 的实验室经过多年研究后，Ilya 形成了一种很少有人认同的直觉：如果神经网络用足够的数据和足够的算力进行训练，它们可以变得比任何人想象的都要强大得多。Ilya 很早就有了这种直觉。

After years of research in Hinton's lab, Ilya had developed an intuition that few people shared. Neural networks could become far more powerful than anyone imagined if they were trained with enough data and enough computing power. Ilya got that intuition very early.

所以，伊利亚一直宣扬只要把规模做大，效果就会更好。我以前总觉得这有点敷衍，认为还需要新想法。结果证明伊利亚基本上是对的，关键在于数据和算力的规模。

So, Ilya was always preaching that you just make it bigger and it'll work better. And I always thought that was a bit of a copout that you're going to have to have new ideas, too. It turns out Ilya was basically right, but it was really the scale of the data and the scale of the computation.

历史上第一次，数据和算力这两样东西大规模地汇聚在一起。原本为游戏设计的强大 GPU 变得可用，使得训练更大的神经网络成为可能。与此同时，互联网带来了数据爆炸。数据集突然变得比研究人员从未见过的还要大，其中有一个数据库脱颖而出——ImageNet。当时最大的图像数据集是 Pascal，它只包含有限数量的类别和相对少量的数据。ImageNet 的规模完全不同：超过 100 万张标注图像，涵盖 1000 个不同类别——狗、汽车、建筑、动物。它不只是另一个数据集，而是有史以来最大的图像数据集。围绕这个数据集有一场比赛，研究团队竞相构建能比任何人都更精确识别图像内容的系统。ImageNet 竞赛是每个严肃 AI 研究者都渴望赢得的终极战场，因为成功意味着证明你的方法能解决现实世界的问题。

For the first time, there was a massive convergence of two things, data and computing power. Powerful GPUs originally designed for gaming were becoming available, making it possible to train much larger neural networks. And at the same time, the internet was creating an explosion of data. Suddenly data sets were becoming bigger than anything researchers had never seen before and among them one database stood above everything else ImageNet. At the time the biggest image data set available was Pascal and Pascal contained only a limited number of categories and a relatively small amount of data. ImageNet was on a completely different scale. More than 1 million labeled images across 1,000 different categories. dogs, cars, buildings, animals. It wasn't just another data set. It was the biggest data set of images ever existed. And around this data set, there was a competition. Teams of researchers were competing to build a system that could recognize what was inside an image more precisely than anyone else. The ImageNet competition was the ultimate battleground that every serious AI researcher wanted to win because success meant demonstrating that your approach could solve real world problems.

我记得有一天伊利亚走进实验室说，看，我们既然让语音识别成功了，这东西真的管用，我们必须赶在别人之前做 ImageNet。

I remember Ilya coming into the lab one day and saying look we now that we got speech recognition working this stuff really works we've got to do ImageNet before anybody else does.

伊利亚立刻看到了大多数研究者忽略的机会。世界第一次拥有了神经网络一直缺失的两样东西：海量数据和前所未有的算力。现在，第一次有可能测试一个激进的想法：直接在原始图像上训练一个巨大的神经网络，让它自己学习。于是他说服实验室的另一位研究员亚历克斯·克里热夫斯基尝试一件几乎没人相信能成功的事：在短短几个月内，用极其有限的计算资源，在有史以来最大的图像数据集上训练一个巨大的神经网络。亚历克斯接受了，并与伊利亚和辛顿一起深入投入这个挑战。截止日期就在那里。2012 年 10 月，意大利佛罗伦萨。比赛日终于到来，ImageNet 的结果即将揭晓。人们就座，饶有兴趣地见证这种不同方法的结果。然后数字出现了：第二名错误率 26.2%，然后是多伦多团队的提交——15.3%。有那么一刻，人们觉得肯定出错了，因为 ImageNet 的改进通常是渐进的，零点几个百分点，也许一个点，而不是 10 个点。他们不只是赢了，而是彻底摧毁了比赛。

Ilya immediately saw an opportunity that most researchers overlooked. For the first time, the world had two ingredients neural networks had always been missing: massive amounts of data and unprecedented computing power. Now, for the first time, it was possible to test a radical idea: train a huge neural network directly on raw images and let it learn on its own. So he convinced Alex Krizhevsky, another researcher in the lab, to attempt something almost nobody believed could work: train a massive neural network on the largest image data set ever assembled in only a few months and with extremely limited computational resources. Alex accepted and with Ilya and Hinton they started working deeply into this challenge. The deadline was there. October 2012, Florence, Italy. The day of competition finally arrived and the ImageNet results were about to be revealed. People took seats to witness with interest the result of this different approach. Then the numbers appeared. Second place, 26.2% error rate. Then supervision, the Toronto team submission, 15.3%. For a moment, people thought something had to be wrong because improvements in ImageNet usually happened gradually, fractions of a percent, one point maybe, not 10. They hadn't won yet destroyed the competition.

对于 ImageNet，所有要素都齐备了。亚历克斯有非常快的卷积核，ImageNet 有足够大的数据，而且确实有机会去做完全前所未有的事情，结果完全成功了。这证明了机器可以自己学习表征。

And with ImageNet, all the ingredients were there. Alex had these very fast convolutional kernels. ImageNet had the large enough data and there was a real opportunity to do something totally unprecedented and it totally worked out. This was the proof machines could learn representations themselves.

几个月内，世界各地的实验室开始放弃传统的计算机视觉流程。到 2014 年，几乎所有人都这样做了。这个领域在一夜之间转向了。

Within months, labs across the world started abandoning traditional computer vision pipelines. By 2014, almost everyone did. The field had pivoted overnight.

你知道吗？这到底是怎么回事？AlexNet 是计算机视觉的突破，还是比那更大的想法？我们得出的结论是，AlexNet 和深度学习所展示的是，现在我们终于有可能——只要有足够的数据、足够的计算规模——就能用计算机解决那些无法用人工设计特征来描述的问题。

You know what? What's going on here? Is this AlexNet a breakthrough in computer vision or is it a bigger idea than that? And we came to the conclusion that what AlexNet and deep learning showed is that it is now finally possible if we had enough data, enough computing scale, we might be able to apply computers to solve problems that were impossible to describe using human engineered features.

AlexNet 之后，几乎立刻有一件事变得清晰：这不只是一个更好的图像识别系统，它看起来像是某种更深刻东西的开端，而谷歌注意到了这一点。到 2013 年，一场新的竞赛已经在硅谷悄然开始。问题不再是深度学习能否工作——AlexNet 已经回答了这个问题。新问题变成了：如果这项技术是真的，谁来控制它？谷歌已经拥有几乎其他公司都没有的东西：数十亿次搜索查询、YouTube、庞大的数据中心。但他们缺少更重要的东西：构建 AI 未来的人。AlexNet 之后仅几个月，谷歌就收购了 DNN Research，这家小公司是围绕杰弗里·辛顿和他的团队创立的。突然之间，伊利亚的工作完全改变了。仅仅几个月前，他还在大学实验室里，资源有限，只有少量 GPU。现在，他拥有了地球上几乎其他研究者都没有的东西——谷歌。

After AlexNet, something became clear almost immediately. This wasn't just a better image recognition system. It looked like the beginning of something much deeper and Google noticed it. Because by 2013, a new race had quietly started inside Silicon Valley. The question was no longer can deep learning work. They had just been answered with AlexNet. The new question became, if this technology is real, who is going to control it? Google already had something almost nobody else had. Billions of search queries, YouTube, massive data centers. But they were missing something far more important. The people building the future of AI. Only a few months after AlexNet, Google acquired DNN Research, the small company created around Geoffrey Hinton and his team. And suddenly, Ilya's work changed completely. Just months earlier, he had been working inside a university lab with limited resources and only a handful of GPUs. Now, he had access to something almost no researchers on Earth had, Google.

在 Google Brain 内部，伊利亚开始与一些最有才华的机器学习研究者共事。但有趣的是，他并没有专注于构建产品，而是专注于一个更大的问题。AlexNet 刚刚证明神经网络可以学习如何“看”，理解图像中的模式。但如果它们能学习的远不止图像呢？因为世界不只是由图片构成的，还有句子中的词语、对话中的声音、问题和答案。与图像不同，词语和语言被认为是一个极其困难的问题，因为词语本身并不携带意义——意义来自词语组合的方式、它们出现的上下文以及使用的顺序。于是伊利亚开始探索一个新想法：神经网络能否像 AlexNet 处理图像那样处理语言？它能否学习隐藏在词语和句子中的模式？例如，将句子从法语翻译成英语，阅读问题并生成答案。换句话说，神经网络能否学会将一个信息序列转换为另一个？这个想法后来被称为序列到序列学习。当时它看起来只是另一篇研究论文，但其中隐藏着更宏大的东西。对伊利亚来说，这引发了一个诱人的可能性：也许神经网络不再局限于识别物体，也许它们也能学会处理语言，而词语几乎可以描述一切。

Inside Google Brain, Ilya started working alongside some of the most talented researchers in machine learning. But interestingly, he wasn't focused on building products. He was focused on a much bigger question. AlexNet had just shown that neural networks could learn how to see, understand patterns inside images. But what if they could learn much more than images? Because the world isn't made only of pictures. It's also made of words in a sentence, sounds in conversations, questions and answers. And unlike images, words and language were considered an incredibly difficult problem because words alone don't carry meaning. Meaning emerges from the way words are combined, the context they appear in, and the sequence in which they are used. So Ilya started exploring a new idea. Could a neural network do with language what AlexNet did with images? Could it learn the patterns hidden inside words and sentences? For example, translate a sentence from French into English, read a question and generate an answer. In other words, could a neural network learn to transform one sequence of information into another? This idea would later become known as sequence-to-sequence learning. At the time, it looked like just another research paper, but hidden inside was something much bigger. For Ilya, this raised an intriguing possibility. Perhaps neural networks weren't limited to recognizing objects anymore. Perhaps they could learn how to work with language as well, and words could describe almost anything.

到 2015 年，伊利亚·苏茨克弗正处于许多 AI 研究者梦寐以求的位置：他在谷歌，这个基本上已成为 AI 宇宙中心的地方。

By 2015, Ilya Sutskever was exactly where many AI researchers dream of being. He was at Google, the place that had basically become the center of the AI universe.

我是谷歌的一名研究员，从事深度学习工作，非常开心。我真的很享受在谷歌的时光。

I was a researcher at Google and I was working on deep learning and I was having a lot of fun. I was really enjoying my time at Google.

他拥有算力，身边有地球上最聪明的一些研究者。所以完全没有理由离开，除了一件事：伊利亚已经开始思考 AI 将走向何方——不只是下一个模型，而是更宏大的东西。

He had computing power. He had some of the smartest researchers on earth around him. So there was absolutely no reason to leave except for one thing. Ilya had begun thinking about where AI was heading. Not just the next model, something much bigger.

然后有一天，想象一下：我正幻想着也许可以创办一家 AI 公司，但完全不清楚该怎么做。你怎么可能弄到那么多钱？这些东西会很贵。

And then one day, picture this: here I am daydreaming that maybe I could start an AI company, but it really wasn't clear how I would do it. How would you possibly get the money for such a thing? Those things will be expensive.

就在这些想法在他脑海中滋长时，一件完全出乎意料的事情发生了。有一天，他收到了一封来自 Sam Altman 的冷邮件，不是什么精心策划的提案，实际上非常简单。

And while these ideas were growing in his mind, something completely unexpected happened. One day an email arrived from Sam Altman, and it wasn't some elaborate proposal. It was actually very simple.

然后有一天，我收到了 Sam 的一封冷邮件，上面写着：‘嘿，来和几个酷人聚聚吧。’我很好奇。

And then actually one day I received a cold email from Sam saying, 'Hey, let's hang out with some cool people.' And I was very curious.

Ilya 去了，坐在桌边的有 Sam Altman、Greg Brockman 和 Elon Musk。当时，Elon Musk 已经因特斯拉和 SpaceX 而声名鹊起，Sam Altman 在运营 Y Combinator，Greg Brockman 刚离开 Stripe。我当然去了，就这样我参加了一场晚餐，他们正在讨论如何创办一个新的 AI 实验室，与当时占据绝对主导地位的 Google 和 DeepMind 竞争。

Ilya went, and sitting there at the table were Sam Altman, Greg Brockman, and Elon Musk. At the time, Elon Musk was already becoming famous for Tesla and SpaceX. Sam Altman was running Y Combinator. Greg Brockman had recently left Stripe. So of course I went, and here I was at a dinner, and they were discussing how could you start a new AI lab which would be a competitor to Google and to DeepMind, which back then had absolute dominance.

与 Google 竞争，与 DeepMind 竞争，在当时听起来很疯狂，就连 Elon Musk 也……

Compete with Google, compete with DeepMind at the time that sounded insane, and even Elon Musk also...

一开始我想，这大概是个无望的努力。我们怎么可能——OpenAI 怎么可能与 Google DeepMind 竞争？我的意思是，这就像蚂蚁对抗大象。

In the beginning I thought, look, this is probably a hopeless endeavor. How could we possibly compete with — how could OpenAI possibly compete with Google DeepMind? I mean, this seemed like an ant against an elephant.

但 Elon Musk 还知道另一件事：如果他们想要哪怕一丝机会，就需要 Ilya，因为他是世界上极少数真正知道如何构建前沿 AI 系统的人之一。他在招募关键科学家和工程师方面发挥了重要作用，最著名的是 Ilya Sutskever。他会说他要加入 OpenAI，然后 Demis 会说服他不加入，然后我又说服他加入。

But Elon Musk also knew something else. If they wanted even a tiny chance, they needed Ilya because he was one of the very few people in the world who actually knew how to build frontier AI systems. He was instrumental in recruiting the key scientists and engineers, most notably Ilya Sutskever. He would say he's going to join OpenAI, then Demis would convince him not to, then I would convince him to do so.

一边是 DeepMind 的 Misosabis，另一边是 Elon Musk——AI 界两个最大的名字试图说服同一个人，因为每个人都明白同一件事：如果 OpenAI 得到了 Ilya，OpenAI 就突然变得真实了。

The Misosabis from DeepMind on one side and Elon Musk on the other — two of the biggest names in AI tried to convince one person, because everyone understood the same thing: if OpenAI got Ilya, OpenAI suddenly became real.

我问 Sam，好吧，我们怎么打破僵局？怎么才能让所有人都说‘好，我们加入’？于是 Altman 想出了一个不寻常的主意：把所有人带出去几天。纳帕谷，没有合同，没有组织架构图，只有研究人员和想法。在那里，有些事情改变了。

I asked Sam, okay, how do we break symmetry here? How do we actually get everyone to say, 'All right, we're joining.' So, Altman came up with an unusual idea. Take everyone away for a few days. Napa Valley, no contracts, no organizational charts, just researchers and ideas. And there something changed.

对吧？想想看，如果我们真的能构建出能够与人类合作解决当今人类面临的最复杂的跨学科挑战的系统，那会怎样？

Right? And you think about what could happen if we could really build systems that are able to work with people on solving the most complicated multi-disciplinary challenges that humanity has to face today.

被核心价值观和使命所吸引，他们中的许多人决定加入。对 Ilya 来说，这意味着做出人生中最大的决定之一。他辞去了 Google 的工作，于 2015 年加入 OpenAI。因为 OpenAI 提供了不同的东西：追求他认为的 AI 的下一个前沿——通用人工智能（AGI）的机会。但在 OpenAI，我们大胆地审视全局。我们问自己，好吧，AI 正走向何方？答案是 AI 正走向 AGI，走向一种最终在各方面与人类一样聪明或更聪明的 AI。那是 OpenAI 最初的愿景：构建 AGI，并且他们想在别人之前做到。但他们也相信同样重要的一点：必须安全地构建。因为如果智能成为人类最强大的技术，那么谁控制了它，谁就能塑造之后的一切。所以从第一天起，OpenAI 就有一个使命：构建 AGI。问题是没人真正知道如何实现。所以在早期，OpenAI 几乎尝试了一切：机器人、游戏、强化学习，通往智能的不同路径，因为如果目标是构建 AGI，路径仍然完全未知。但当许多研究人员探索不同方向时，Ilya 不断回到一个想法：也许智能不是你能手动构建的。也许智能可以涌现。

Engaged by the core values and missions, many of them decided to come on board. For this meant making one of the biggest decisions of his life. He resigned from Google and joined OpenAI in 2015. But because OpenAI was offering something different: the opportunity to pursue what he believed was the next frontier of artificial intelligence — artificial general intelligence. But at OpenAI, we took the liberty to look at the big picture. We asked ourselves, okay, what's the where is AI going towards? And the answer is AI is going towards AGI, towards an AI which eventually is as smart or smarter than a human in every way. That was the original vision of OpenAI: build AGI, and they wanted to build it before anyone else did. But they also believed something equally important: that it had to be built safely. Because if intelligence became humanity's most powerful technology, whoever controlled it could shape everything that came after. So starting from day zero, OpenAI had one mission: build AGI. The problem was nobody actually knew how to get there. So during the early years, OpenAI tried almost everything: robotics, games, reinforcement learning, different paths toward intelligence, because if the goal was building AGI, the path was still completely unknown. But while many researchers were exploring different directions, Ilya kept returning to one idea: maybe intelligence wasn't something you manually build. Maybe intelligence can emerge.

就在 OpenAI 成立之前，我遇到了 Ilya。他对我说的第一件事是：‘听着，模型们只是想学习。你必须明白这一点。模型们只是想学习。你把障碍从它们面前移开，对吧？你给它们好的数据。你给它们足够的操作空间。你不要做蠢事，比如在数值上糟糕地调节它们。它们想学习。它们会做到的。它们会做到的。’

Just before OpenAI started, I met Ilya. One of the first things he said to me was, 'Look, the models, they just want to learn. You have to understand this. The models, they just want to learn. You get the obstacles out of their way, right? You give them good data. You give them enough space to operate in. You don't do something stupid like condition them badly numerically. And they want to learn. They'll do it. They'll do it.'

起初，这听起来几乎像哲学。‘模型们只是想学习。’但这句话背后隐藏着一个激进的想法。想象一下教一个孩子。你不会解释他们生活中将面临的每一种可能情况。你不会写下成千上万条规则。相反，你让他们接触书籍、对话、经历、模式，慢慢地他们开始理解世界。所以 Ilya 开始问一个奇怪的问题：如果我们对 AI 做同样的事情呢？如果你只给它一个目标：预测接下来会发生什么？Ilya Sutskever 曾经说过一句非常简单的话，一直印在我脑海里：‘预测非常接近智能。’这个想法是，如果你能把关于世界的所有信息、事物的状态等等压缩成最小的表示，然后作为其中的一部分预测接下来会发生什么，你就以一种深刻的方式理解了它，因为要极其准确地预测下一个词，你不能简单地记忆单词。你需要理解事实、关系、模式、人类行为、世界本身如何运作——这是对人类智能的最佳近似。而在 OpenAI 内部，这个想法开始变得越来越重要。所以他们开始将这个概念应用到真实模型上，结果很有希望。他们一次又一次地改进。这不再是随机的了。每次他们增加更多的算力、更多的数据、更大的模型，系统都在不断改进。所以你会说，嘿，如果你把一些算力和一些数据混合到一个特定规模的神经网络中，你会得到结果，而且你知道如果你只是把配方放大，结果会更好。起初是实验的东西变成了一种策略。Scaling——一个词，但这个词开始改变整个 AI 行业。

At first, this almost sounds philosophical. 'The models just want to learn.' But hidden inside that sentence was a radical idea. Imagine teaching a child. You don't explain every possible situation they will face in life. You don't write thousands of rules. Instead, you expose them to books, conversations, experiences, patterns, and slowly they begin understanding the world. So Ilya started asking a strange question: what if we did the same thing with AI? What if you gave it only one objective: predict what comes next? Ilya Sutskever once said a very simple sentence that stuck in my mind: 'Prediction is very close to intelligence.' And the idea is that if you can compress all of the information about the world, the state of things, whatever, into its smallest representation, and then as part of that predict the thing that's going to happen next, you understand it in a sort of deep way, because to predict the next word extremely well, you can't simply memorize words. You need to understand facts, relationships, patterns, human behavior, how the world itself works — the best approximation possible of human intelligence. And yet inside OpenAI, this idea started becoming increasingly important. So they started applying this concept to real models, and the results were promising. They kept improving again and again. This wasn't random anymore. Every time they increased more compute, more data, larger models, the systems kept improving. So you say, hey, if you mix some compute with some data into a neural net of a certain size, you will get results, and you will know that you will be better if you just scale the recipe up. What started as an experiment became a strategy. Scaling — one word, but a word that started changing the entire AI industry.

然后 Scaling 的洞察出现了，对吧？缩放定律，GPT-3，突然每个人都意识到我们应该 Scaling。这只是一个例子，说明语言如何影响思想。Scaling 只是一个词，但它如此强大，因为它告诉人们该做什么。做这个。好吧，让我们尝试 Scaling 事物。

Then the scaling insight arrived, right? Scaling laws, GPT-3, and suddenly everyone realized we should scale. And it's just — this is an example of how language affects thought. Scaling is just one word, but it's such a powerful word because it informs people what to do. Do this. Okay, let's try to scale things.

很快，一场竞赛在硅谷和整个研究界开始了。每个人都突然想要更多的数据、更多的 GPU、更大的模型。与此同时，OpenAI 继续前进。GPT-1、GPT-2、GPT-3。

And very quickly, a race started across Silicon Valley and the entire research community. Everyone suddenly wanted more data, more GPUs, bigger models. And in the meanwhile, OpenAI kept moving forward. GPT-1, GPT-2, GPT-3.

每个模型都比前一个更大，每个模型都在更多数据上训练，每个模型都变得越来越强大。但随后，一些意想不到的事情开始发生。系统不仅仅是变得更好，它们开始做一些没人明确教过它们的事情：写故事、总结信息、生成代码、回答问题，仿佛新能力是自然涌现的。就在那一刻，OpenAI 内部的人开始意识到，这已经不再正常了。他们创造的东西远比大多数人预期的要强大。到 2022 年底，他们面临一个非常不寻常的决定。多年来，他们的模型大多停留在研究论文和内部实验中。现在，情况不同了：一个人们可以实际使用的模型诞生了。但问题在于，没人真正知道如果把这样一个系统直接交给公众会发生什么。技术改进得太快，连构建它的人都不完全清楚它会走向何方。

Each model larger than the previous one, each one trained on more data, each one becoming increasingly capable. But then something unexpected started happening. The systems weren't simply becoming better. They started doing things nobody explicitly told them. Writing stories, summarizing information, generating code, answering questions, almost as if new abilities were appearing naturally on their own. And that was the moment people inside OpenAI started realizing this wasn't normal anymore. What they created was way more capable than what most expected. And by late 2022, they faced a very unusual decision. For years, their models had lived mostly inside research papers and internal experiments. Now, something different: a model that people could actually use. But there was a problem. Nobody really knew what would happen if they put a system like this directly into the hands of the public. The technology was improving so quickly that even the people building it didn't fully understand where it might lead.

但一旦它变得非常复杂，我们实际上并不比了解你的大脑更了解它内部发生了什么。

But as soon as it gets really complicated, we don't actually know what's going on any more than we know what's going on in your brain.

你说我们不知道它具体怎么工作是什么意思？它可是人设计的。

What do you mean we don't know exactly how it works? It was designed by people.

不，不是的。我们做的是设计了学习算法，这有点像设计进化原理。但当这个学习算法与数据交互时，它会产生复杂的神经网络，这些网络擅长做事，但我们并不真正理解它们具体是如何做到的。

No, it wasn't. What we did was we designed the learning algorithm. That's a bit like designing the principle of evolution. But when this learning algorithm then interacts with data, it produces complicated neural networks that are good at doing things, but we don't really understand exactly how they do those things.

尽管如此，OpenAI 内部又出现了另一种信念：你无法通过把如此强大的系统锁在实验室里来理解它，你必须看到真实世界会如何与之互动。因此，他们没有等待完美，而是决定以研究预览的形式发布它。不是给研究人员或科学家，而是给所有人。2022 年 11 月 30 日，他们按下了按钮。

Still, inside OpenAI, there was another belief emerging. You couldn't understand a system this powerful by keeping it locked inside a laboratory. You had to see how the real world would interact with it. So rather than waiting for perfection, they decided to release it as a research preview. Not to researchers or scientists, to everyone. And on November 30, 2022, they pressed the button.

非常诡异。一款新的人工智能工具因能在几秒钟内写出整篇文章而走红，它叫 ChatGPT。5 天内，就有 100 万用户访问 ChatGPT。两个月内，用户数达到 1 亿，成为历史上增长最快的消费级应用。学生、作家、程序员、公司，每个人都在尝试它。互联网开始疯狂。ChatGPT，也许你听说过它。如果你还没听过，那请做好准备，因为它有望成为彻底改变我们做事方式的病毒式轰动。伊利亚花了多年时间帮助构建的东西，现在正成为地球上最具影响力的技术之一。到 2023 年底，OpenAI 已成为地球上最重要的公司之一。不到一年前，ChatGPT 几乎还不存在。而现在，数亿人在使用 AI。这一切的中心是 Sam Altman。在公众眼中，他已成为人工智能的代言人，带领人类进入下一个时代的人。但在 OpenAI 内部，情况正在发生变化。因为尽管公司发展速度超乎想象，但关于安全的争论并未消失，反而变得更加激烈，而没有人比 Ilya Sutskever 更深入地思考这些问题。

Very creepy. A new artificial intelligence tool is going viral for cranking out entire essays in a matter of seconds and it's called Chat GPT. Within 5 days, 1 million users were accessing ChatGPT. Within two months, 100 million users. The fastest growing consumer application in history. Students, writers, programmers, companies, everyone was trying it. The internet started losing its mind. Chat GPT. Maybe you've heard of it. If you haven't, then get ready because this promises to be the viral sensation that could completely reset how we do things. The thing Ilya had spent years helping build was now becoming one of the most impactful technologies on planet Earth. By the end of 2023, OpenAI had become one of the most important companies on Earth. Less than a year earlier, ChatGPT had barely existed. Now, hundreds of millions of people were using AI. And at the center of it all was Sam Altman. To the public, he had become the face of artificial intelligence, the man leading humanity into the next era. But inside OpenAI, something was changing. Because while the company was growing faster than anyone imagined, the debates around safety weren't disappearing, they were becoming more intense and nobody was thinking more deeply about those questions than Ilya Sutskever.

我确实认为，我非常喜欢 Ilya 的一点是，他非常认真地对待 AGI 和广泛的安全问题，包括这对社会的影响。过去几年里，Ilya 是我花最多时间讨论这将意味着什么的人之一。

I do think one of the many things that I really love about Ilya is he takes AGI and the safety concerns broadly speaking, you know, including things like the impact this is going to have on society very seriously. Ilya is one of the people that I've spent the most time over the last couple of years talking about what this is going to mean.

多年来，Sam 和 Ilya 并肩工作。一个推动组织前进，另一个帮助定义背后的科学。他们一起将 OpenAI 从一个小型非营利组织转变为世界上最重要的 AI 公司。但现在，曾经让他们走到一起的成功似乎正在将他们分开。2023 年 11 月 17 日。表面上看，这似乎是平常的一天。但在紧闭的门后，董事会已经进行了数周的谈话。这些谈话无人知晓——员工不知道，投资者不知道，甚至微软也不知道。然后突然间，一切都变了。科技界在周末陷入混乱，因为给我们带来 ChatGPT 的公司解雇了其 CEO Sam Altman，这位被比作史蒂夫·乔布斯等科技巨头的 CEO 于周五被 OpenAI 董事会解雇。

For years, Sam and Ilya had worked side by side. One pushing the organization forward, the other helping define the science behind it. Together, they had transformed OpenAI from a small nonprofit into the most important AI company in the world. But now, the same success that brought them together seemed to be pulling them apart. November 17th, 2023. Apparently a day like another, at least from the outside. Behind closed doors, the board had been having conversations for weeks. Conversations nobody knew about. Not employees, not investors, not even Microsoft. Then suddenly everything changed. The tech world has been thrown into chaos over the weekend when the company that gave us ChatGPT fired its CEO Sam Altman who has drawn comparisons to tech giants like Steve Jobs was dismissed by the OpenAI board Friday.

整个硅谷都震惊了。员工和投资者都震惊了。即使是最接近 OpenAI 的人也难以理解发生了什么。而支持这一决定的人中，就有 Ilya Sutskever 本人。

The entire Silicon Valley was stunned. Employees and investors were stunned. Even the people closest to OpenAI struggled to understand what had happened. And among the people who supported the decision was Ilya Sutskever himself.

我特别自豪的是，我的一个学生解雇了 Sam——那个与 Sam 一起花了近十年建立 OpenAI 的人。没人知道该怎么理解这件事。Ilya 为什么要这么做？有什么理由能证明罢免世界上增长最快的科技公司的 CEO 是合理的？各种猜测立刻涌现：安全、权力、治理、AGI，没人知道。

I'm particularly proud of the fact that one of my students fired Sam, the same person who had spent nearly a decade building OpenAI alongside Sam. Nobody knew what to make of it. Why would Ilya do this? What could possibly justify removing the CEO of the fastest growing technology company in the world? The theories started immediately: safety, power, governance, AGI, nobody knew.

我认为我们应该对此感到担忧，因为我认为 Ilya 实际上有很强的道德指南针。他思考，他非常纠结于什么是对的问题。如果他强烈到想要解雇 Sam，那么我认为世界应该知道那个原因是什么。

I think we should be concerned about this because I think Ilya actually has a strong moral compass. He thinks about, he really sweats it over questions of what is right. And if he felt strongly enough to want to fire Sam, well, I think the world should know what was that reason.

就连 Elon Musk 也在寻找答案。OpenAI 发现了什么？Ilya 看到了什么让他担心的事情？要么这是一件严肃的事情，我们应该知道它是什么；要么这不是一件严肃的事情，那么董事会就应该辞职。

Even Elon Musk was searching for answers. Had OpenAI discovered something? Had Ilya seen something that worried him? Either it was a serious thing and we should know what it is or it was not a serious thing and then the board should resign.

直到今天，Ilya Sutskever 从未公开解释他为何投票罢免 Sam Altman。官方解释集中在治理以及董事会与 Altman 之间信任破裂上。但 AI 社区内部的许多人怀疑有更深层的原因：关于 OpenAI 是否发展过快的问题。

To this day, Ilya Sutskever has never publicly explained why he voted to remove Sam Altman. The official explanation focused on governance and a breakdown of trust between the board and Altman. But many people inside the AI community suspected something deeper: questions about whether OpenAI was moving too fast.

所以，OpenAI 成立时非常强调安全。其主要目标是开发通用人工智能并确保其安全。我的一位前学生 Ilya Sutskever 是首席科学家，但随着时间的推移，事实证明 Sam Altman 对安全的关注远不如对利润的关注，我认为这很不幸。

So, OpenAI was set up with a big emphasis on safety. Its primary objective was to develop artificial general intelligence and ensure that it was safe. One of my former students, Ilya Sutskever, was the chief scientist and over time it turned out that Sam Altman was much less concerned with safety than with profits and I think that's unfortunate.

在混乱的几天里，OpenAI 陷入了危机。然后，意想不到的事情再次发生。Ilya 改变了主意。Sam Altman 回归，公司继续前进——至少表面上是这样，因为仅仅几个月后，Ilya 本人就离开了 OpenAI。当他最终透露他接下来想构建什么时，很明显，真正的故事从来不是关于 Sam Altman 的。而是关于更宏大的事情：人工智能可能真的会成功。OpenAI 危机几个月后，Ilya Sutskever 离开了。对许多人来说，这说不通。OpenAI 赢了。ChatGPT 征服了世界。公司价值数十亿美元。

For a few chaotic days, OpenAI descended into crisis. Then something unexpected happened again. Ilya changed his mind. Sam Altman returned and the company moved forward, at least on the surface, because only a few months later he himself would leave OpenAI. And when he finally revealed what he wanted to build next, it became clear that the real story was never about Sam Altman. It was about something much bigger. The possibility that artificial intelligence might actually work. A few months after the OpenAI crisis, Ilya Sutskever left. For many people, it didn't make sense. OpenAI had won. ChatGPT had conquered the world. The company was worth billions.

那么，他为什么要离开？要理解这一点，我们需要了解伊利亚认为接下来会发生什么。慢慢地但肯定地，或者也许没那么慢，AI 会不断变得更好。总有一天，AI 会做到我们能做的所有事情。不是一部分，而是全部。

So why leave? To understand that, we need to understand what Ilia believed was coming next. Slowly but surely, or maybe not so slowly, AI will keep getting better. And the day will come when AI will do all the things that we can do. Not just some of them, but all of them.

对大多数人来说，AI 是一个聊天机器人、一个生产力工具、一个更好的搜索引擎。但伊利亚想的不是聊天机器人。他在想之后会发生什么。因为如果智能可以 Scaling（规模扩张），那么最终这个系统可能能够执行人类能完成的每一项智力任务。一旦你相信这一点，一个完全不同的问题就出现了。AI 和 AGI 的问题是什么？整个问题就在于权力。当权力变得非常大时，会发生什么？

For most people, AI was a chatbot, a productivity tool, a better search engine. But Ilia wasn't thinking about chatbot. He was thinking about what comes after. Because if intelligence can be scaled, then eventually this system may become capable of performing every intellectual task a human can perform. And once you believe that, a completely different question appears. What is the problem of AI and AGI? The whole problem is the power. When the power is really big, what's going to happen?

担忧在于，当智能变得强大时会发生什么，因为技术本身可能第一次能够做出决策，甚至构建它的人也承认一些令人不安的事情：他们并不完全理解它是如何工作的。

The concern was what happened when intelligence becomes powerful because for the first time the technology itself may become capable of making decisions and even the people building it admit something uncomfortable. They don't fully understand how it works.

还有一个长期的生存威胁，当我们创造出比我们更聪明的数字生命时，这个威胁就会出现。我们不知道我们是否能保持控制。几十年来，这听起来像科幻小说。现在，它正被创造这项技术的极少数人讨论。

There is also a longer-term existential threat that will arise when we create digital beings that are more intelligent than ourselves. We have no idea whether we can stay in control. For decades, this sounded like science fiction. Now it was being discussed by the very few people who created the technology.

根据辛顿的说法，还有另一个担忧的理由。但我们有证据表明，如果它们是由追求短期利润的公司创造的，我们的安全就不会是首要任务。

And according to Hinton, there was another reason to worry. But we now have evidence that if they are created by companies motivated by short-term profits, our safety will not be the top priority.

这个想法变得越来越难以忽视，因为在 ChatGPT 之后，AI 不再只是一个科学项目。它变成了一场竞赛。公司筹集了数十亿美元。激励机制正在改变，伊利亚相信还会发生别的事情。

This idea was becoming increasingly difficult to ignore because after ChatGPT, AI wasn't just a scientific project anymore. It had become a race. Companies were raising billions. The incentives were changing and Ilia believed something else would happen.

我还认为，随着 AI 变得越来越强大，越来越明显地强大，政府和公众也会产生采取行动的愿望。我确实认为，在某个时候，AI 会开始让人感觉真的很强大。我认为当这种情况发生时，我们会看到所有 AI 公司对待安全的方式发生巨大变化。他们会变得更加偏执。

I also maintain that as AI continues to become more powerful, more visibly powerful, there will also be a desire from governments and the public to do something. I do think that at some point the AI will start to feel powerful actually. And I think when that happens, we will see a big change in the way all AI companies approach safety. They'll become much more paranoid.

换句话说，现在人们仍然关注 AI 出错的地方：错误、幻觉、失败。但伊利亚相信，有一天这种讨论会改变。如果超级智能无论如何都会到来，那么就需要有人完全专注于让它安全。于是在 2024 年 6 月，他重新开始了。

In other words, right now people still focus on what AI gets wrong, the mistakes, the hallucinations, the failures. But Ilia believed that one day the conversation would change. If superintelligence was coming anyway, someone needed to focus entirely on making it safe. So in June 2024, he started over.

又一家全新 AI 初创公司获得了令人瞠目的估值。这一次是 OpenAI 的前首席科学家为自己的公司 Safe Superintelligence, Inc. 筹集了高达十亿美元的资金。

Another eye-popping valuation for a brand new AI startup. This time it's OpenAI's former chief scientist raising a whopping billion dollars for his own company, Safe Superintelligence, Inc.

离开 OpenAI 几周后，伊利亚透露了他一直在做的事情：一家名为 Safe Superintelligence（简称 SSI）的新公司。他与成功的创业者和 AI 投资者 Daniel Gross，以及之前与他一起在 OpenAI 工作的研究员 Daniel Levi 共同创立了这家公司。

A few weeks after leaving OpenAI, Ilia revealed what he had been working on, a new company called Safe Superintelligence, or simply SSI. He founded it together with Daniel Gross, a successful entrepreneur and AI investor, and Daniel Levi, a researcher who had previously worked alongside him at OpenAI.

但让 SSI 与众不同的不是团队，而是使命。因为与 OpenAI、Anthropic 以及世界上几乎所有主要 AI 公司不同，SSI 并不试图构建一个聊天机器人。事实上，该公司明确表示，他们无意在研究压力与商业压力之间取得平衡。没有消费者应用，没有用户争夺战，没有产品路线图，只有一个目标：构建超级智能并确保其安全。

But what made SSI different wasn't the team, it was the mission. Because unlike OpenAI, unlike Anthropic, and unlike virtually every major AI company in the world, SSI wasn't trying to build a chatbot. In fact, the company explicitly stated that they had no intention of balancing research with commercial pressures. No consumer applications, no race for users, no product roadmap, only one objective: build a superintelligence and make sure it is safe.

对伊利亚来说，这种区别至关重要，因为他相信，一旦一家公司开始争夺客户、收入和增长，它就不可避免地面临压力：要更快行动、发布产品、优先考虑短期目标。SSI 的设计就是为了完全避免这种冲突。它的唯一焦点就是研究，别无其他。

For Ilia, that distinction was crucial because he believed that once a company starts competing for customers, revenue, and growth, it inevitably faces pressure to move faster, to release products, to prioritize short-term goals. SSI was designed to avoid that conflict entirely. Its singular focus would be research, nothing else.

投资者立即注意到了。尽管没有公开产品、没有收入、只有 10 名员工，SSI 还是迅速达到了数十亿美元的估值。不是因为已经构建了什么，而是因为谁在构建它。许多投资者相信，如果有人理解人工智能的发展方向，那一定是当初帮助创建 ChatGPT 的人之一。

Investors immediately paid attention. Despite having no public product, no revenue, and just 10 employees, SSI quickly reached a multi-billion dollar valuation. Not because of what it had built, but because of who was building it. Many investors believe that if anyone understood where artificial intelligence was heading, it was one of the people who helped create ChatGPT in the first place.

SSI 已经筹集了 30 亿美元。但你可以说，看看其他公司筹集了更多。但他们的大量算力都用于推理。这些大数字、这些大额贷款，都是指定用于推理的。如果你想要一个进行推理的产品，你需要大量的工程师和销售人员。很多研究需要专门用于生产各种产品相关功能。所以当你看看实际留给研究的算力时，差距就小得多了。我们有足够的算力来证明并说服我们自己和其他人，我们正在做的事情是正确的。

SSI has raised $3 billion. But you could say, but look at the other companies raising much more. But a lot of what their compute goes for inference. Like these big numbers, these big loans, it's earmarked for inference. You need if you want to have a product on which you do inference, you need to have a big staff of engineers, of salespeople. A lot of the research needs to be dedicated for producing all kinds of product related features. So then when you look at what's actually left for research, the difference becomes a lot smaller. We have sufficient compute to prove to convince ourselves and anyone else that what we're doing is correct.

像 OpenAI、Google 或 Anthropic 这样的公司必须将资源分配给研究、产品、基础设施和服务数百万用户。SSI 只有一个目标：研究。每一美元、每一个 GPU、每一位研究人员都专注于一个单一问题：如何构建安全的超级智能。

Companies like OpenAI, Google or Anthropic must divide their resources between research, products, infrastructure and serving millions of users. SSI has only one goal: research. Every dollar, every GPU, every researcher focuses on a single problem: how to build a safe superintelligence.

在伊利亚看来，这种专注让 SSI 尽管比竞争对手小得多，但仍能保持竞争力，因为他们不是试图做所有事情。他们试图把一个问题解决得异常出色。

In Ilia's view, that focus allows SSI to remain competitive despite being much smaller than its rivals because they're not trying to do everything. They are trying to solve one problem exceptionally well.

伊利亚·苏茨克弗的故事始于一个对神经网络着迷的学生。他帮助创造了 AlexNet、ChatGPT 和现代 AI 革命。但今天，他的使命不同了：不是让 AI 更强大，而是确保当它变得足够强大时，人类仍然保持控制。因为也许人工智能中最重要的问题不再是“我们如何构建它”，而是“如果它真的成功了会发生什么”。

The story of Ilia Sutskever began with a student fascinated by neural networks. He helped create AlexNet, ChatGPT and the modern AI revolution. But today his mission is different: not making AI more powerful, but making sure that when it becomes powerful enough, humanity remains in control. Because perhaps the most important question in artificial intelligence is no longer how do we build it, but what happens if it actually works.

至此，我们来到了本期视频的结尾。如果你觉得有趣，请点赞并订阅频道。下次见，Vision Economy。

And with that, we have reached the end of this video. If you found it interesting, leave a like and subscribe to the channel. Until next time from Vision Economy.