Alexander Wang：数据是新代码，以及 AI 的未来

欢迎收听《Logan Bartlett 秀》。本期节目，你将听到我与亚历山大·王的对话。亚历山大是 Scale 的联合创始人兼 CEO，该公司最新估值 70 亿美元，帮助企业利用数据开发人工智能模型。亚历山大 19 岁辍学创办这家公司，如今它已成长为人工智能领域最重要的公司之一。我们聊了 AI 的未来，包括灾难性风险，以及他对 AI 可能加剧社会不平等的担忧。还谈到了运营经验，比如招聘真正在乎问题的人。和这位全球最年轻的亿万富翁之一聊得很开心。亚历克斯，感谢你的到来。

Welcome to the Logan Bartlett Show. On this episode, what you're going to hear is a conversation I have with Alexander Wang. Now, Alexander is the co-founder and CEO of Scale, a company most recently valued at $7 billion that helps companies use their data as an input into the development of artificial intelligence models. Alexander started this company at 19 after dropping out of school and it scaled into one of the most important companies in the world of artificial intelligence today. really interesting conversation with Alexander about the future of artificial intelligence, including what the risk of catastrophic doom is, as well as his concerns about the potential for artificial intelligence to create further inequality in society. We also talk about his operational lessons, including hiring people that actually give a about the problems you're solving. Really fun conversation with one of the world's youngest billionaires in Alexander that you'll hear now. Alex, thanks for doing this.

当然，谢谢邀请。

Of course, thanks for having me.

大约十年前，有个说法很流行：数据是新的石油。你为什么反对这个观点？

So, there was a phrase that was pretty ubiquitous about a decade ago that data was the new oil. Can you talk about why you reject that view?

我觉得这个说法有很多对的地方。打个比方，二十年前，全球最大的公司都是石油公司。那时石油是权力和杠杆的源泉，主要是经济杠杆。所以数据是新的石油，意思是在未来几十年，它将成为经济力量和经济影响力的主要杠杆。但错的地方在于，数据不是一种商品。并非所有数据生而平等，不像石油那样。石油是稀缺商品，但数据丰富得多。数据多种多样，有代码数据、语言数据、法律数据，每种都不同。因此，从战略上看，你需要不同的框架。你不能到处找数据井，挖出来卖掉。你需要深思熟虑的策略，把不同质量的数据源拼接起来。

So, I think there's a lot that the phrase gets right. I think that's sort of like one framing is so if you went back like two decades, the largest companies in the world were all oil companies. And so at that point, and less the case now, but oil and petroleum were sort of like the bringers of power and leverage, mostly economic leverage. So I think the way in which data is the new oil is that it is by and large going to be the main lever for economic power and economic influence over the course of the next few decades. I think the thing that it gets wrong is that data is not a commodity in the same way. It's not like not all data is created equal in the same way as oil. So you know oil by definition is like a scarce commodity. But data is far richer than that. Data has multitudes. You could have data specific to code or data specific to language or data specific to law. And each of these pieces of data is quite different. And therefore, when you think about it strategically, it's a different framework you have to apply. You're not just going around hunting for data wells and just try to mine them up and resell them. You need a thoughtful strategy by which you're stitching together useful qualitatively different data sources.

那么“数据是新的代码”是什么意思？它如何成为 Scale 创立的基础？

What does data is the new code mean and how did that serve as a primitive to the founding of scale?

基本概念是：什么构建块能支撑下一代应用？过去 50 年，毫无疑问是代码。代码带来了许多技术革命，尤其是互联网和移动。代码是基础构建块。但展望 AI 时代，模型和算法越来越成为我们交互、驱动应用的核心，数据就成了构建块。我的亲身经历是，在 MIT 上学时，谷歌发布了 TensorFlow。那是深度学习和大型神经网络开始普及的早期。我记得用完全相同的算法检测面部情绪，也用来检测冰箱里的食物是否不见了。什么都没变，只有数据变了。代码一样，算法一样，终端上跑的命令也一样，只是数据改变了算法性能。所以，如果你思考未来 50 年的技术，什么能让一个应用与众不同？什么构建块组合起来能创造出非凡的东西？那就是数据。这就是核心洞察。

The basic concept is that what is the building block that enables the next generation of applications? And I think that building block undeniably for the past 50 years has been code. Code has enabled many revolutions in technology, most notably the internet and mobile. And code was that fundamental building block. I think as you peer forward towards the era of AI, in a world where models and algorithms more and more start to be what we interact with, govern the applications we use, like be the core primitive of our technological lives, then data actually becomes the building block. The formative experience for me here was I was in college at MIT right when Google released TensorFlow. It was the very early moments where deep learning and large neural networks were starting to become democratized. I remember using the exact same algorithm to detect facial emotions as to detect whether or not my food had gone missing inside my fridge. Nothing had changed, just data. The code was all the same. The algorithms were all the same. You run the exact same commands on the terminal and it was just data changing the performance of the algorithm. So the formative experience was basically if you think about the next 50 years of technology, what is going to differentiate one application from another? What are those building blocks you're going to compose on top of one another that's going to make an incredibly differentiated thing or something that delights consumers? And that thing was data, which gets at the heart of its importance going forward. So that's the insight.

能举个早期用例的例子吗？是什么让你开始做这个的？

Can we walk through a specific example of a use case in the early days that kind of got you going around this?

最早的用例全是自动驾驶汽车。回到 2016、2017 年的硅谷，大趋势是自动驾驶。很多公司成立，车企也启动项目。通用收购 Cruise 可能打响了发令枪。所有自动驾驶汽车都有一个要求：完全看清路上的一切。车能开在路上，看到“那里有个人，有辆车，有个骑自行车的人，有个施工锥。交通灯是这样显示的。”它们需要完全理解周围环境。为此，它们必须构建算法，摄入海量数据，成千上万的例子让算法学习：这个场景下，车都在这里；那个场景下，人都在这里；这个场景下，行人都在这里。然后基于数百万个这样的例子训练，打造稳健的车辆。现在算是循环了，旧金山到处是无人驾驶汽车，终于成为现实。

Yeah, the earliest use case was all autonomous vehicles. So go back to 2016, 2017 in Silicon Valley, the mega trend was autonomous vehicles and self-driving. There were many companies being started. A lot of the automakers were starting their own programs. There was the GM Cruise acquisition, which was maybe the starting gun for the entire industry. All these autonomous vehicles had one requirement to be self-driving: they need to fully see everything on the road. These cars can drive down the road and can see 'oh there's a person there, there's a car there, there's a bicyclist there, there's a construction cone over there. This is what the traffic light says.' They need to fully understand the environment around them. To do that, they had to build algorithms that ingested huge amounts of data, tons of examples where the algorithm could learn from: in this scenario, this is where all the cars were; in this scenario, this is where all the people were; this scenario, where all the pedestrians were. And then train off millions of examples like that to build these robust vehicles. It's kind of come full circle now because in San Francisco you have self-driving cars driving around everywhere without drivers, and it's finally become a reality.

Scale 在自动驾驶的价值链中扮演什么角色？你们在哪儿切入，和 Cruise、Waymo 这些公司有什么不同？

What did Scale play in that value chain of getting autonomous cars going? Where did you fit in versus where Cruise stopped or Waymo or whatever the right example is?

具体是在数据精炼阶段。车辆收集大量数据，到处开，得到大量影像、视频、激光雷达数据、雷达数据，所有传感器数据。

It was specifically in this data refinement stage where the cars would collect huge amounts of data. They would drive around, you would get tons of footage, video footage, lidar data, radar data, all this sensor data altogether.

但在那些数据中，没有任何实际标注的例子，比如这是人的位置、这是行人的位置、这是自行车的位置、这是汽车的位置。所以算法没有东西可以学习。我们做的是将原始数据转化为所谓的标注数据或高质量数据，用于机器学习应用，所有这些例子都被标记出来，这样模型就能真正学习在什么情况下人长什么样、行人长什么样、汽车长什么样等等。我们喜欢说的一句话是，虽然我不同意‘数据是新的石油’这种说法，但如果数据是新的石油，那么规模就是炼油厂。我们经历了这样一个过程：将大量原始数据转化为非常高质量的数据，然后为你的算法提供动力。

But in none of that data were there actual examples marked of this is where a person is, this is where a pedestrian is, this is where a bicyclist is, this is where a car is. And so the algorithm had nothing to learn off of. So what we did is we went from raw data to what's called labeled data or high quality data for machine learning applications where all these examples were marked so that the model could actually learn in what situations what does a person look like, what does a pedestrian look like, what does a car look like, etc. One of the things that we like to say, while I disagree with the framing that data is the new oil, if data is the new oil then scale is the refinery. And we underwent this process by which you would convert large amounts of raw data to very high quality data that can then power your algorithms.

那为什么他们想把这个问题外包给第三方，而不是自己内部建立这种能力呢？

And why was that a problem that they wanted to outsource to a third party rather than bringing that in-house and building that competency out themselves?

我认为总的来说，如果你看看整个 AI 行业，那些大规模的基础构件或大规模要素最终都是如此巨大的问题，以至于值得成立公司来占据这些基础设施的位置。另一种思考方式是，当我创办 Scale 时，深受 Stripe 和 AWS 的启发，这些大规模基础设施公司非常有远见，因为它们基本上意识到每个行业、每个初创公司都会面临同样的问题。它们把这些东西拿来，为开发者打造了几乎像消费级一样的体验，并且把它们做到非常易用，规模经济效应如此明显，以至于它们成为了行业默认选择。所以如果你看看 AI 或机器学习，主要有三个要素：算力，即 GPU 和其他芯片，用于驱动极其数据密集和计算密集的算法，正如我们所见，现在几乎整个行业都外包给 Nvidia；人才，这实际上无法外包，但人才是这些公司显然花费巨额资金的地方。这些公司的工程师赚着数百万美元。他们有数百人的团队。所以他们在人才上花费了数十亿美元。然后是数据。这三个要素都是整个 AI 组件树中如此大的部分，以至于如果有公司能够以非常高水准和高质量的方式解决它们，它们就会被当作基础设施使用。行业需要为每个组件提供基础设施层。这就是我看待它的方式。对于每个公司来说，它们都有这个选择：是自己内部构建还是使用行业基础设施？大多数公司采取的方法是：有一些事情自己构建以形成差异化是有意义的，但你必须接受，由于基础设施提供商拥有的规模经济和网络效应，你通常会比行业标准效率更低。

I think in general if you look at the overall AI industry, the large-scale building blocks or the large scale ingredients for it end up being such big problems that companies deserve to be built to occupy those infrastructure slots. Another way to think about this is like when I was starting Scale, I was very inspired by Stripe and AWS, these large scale infrastructure companies that felt very visionary because they basically realized that there were the same problems that every company in a sector, every company in the startup industry were going to deal with. And they basically took those and built almost like consumer level experiences for the developers and built them to a point where it's so easy to use and the economies of scale were so clear that they just became the defaults within the industry. So if you look at that for AI or for machine learning, there were kind of three main ingredients: there's compute, so GPUs and other chips to power the incredibly data intensive and compute intensive algorithms, and as we've seen almost the entire industry outsources to Nvidia at this point; there's talent, which there's no way to outsource really, but talent is this place where these companies obviously are spending huge amounts of money. Engineers at these firms are making millions and millions of dollars. They have teams of hundreds and hundreds of people. So they're spending on the order of billions of dollars on talent full stop. And then there's data. Each of these three ingredients were such big pieces of the overall AI component tree that if there were companies that could solve them in a very high class way and a very high quality way, they were going to be used like infrastructure. The industry demanded an infrastructure layer at for each of these components. So that's really the way I look at it. For each individual company they have this option: do I build it in-house or do I use the industry infrastructure? Most companies take an approach which is like there are a few things where it makes sense for me to build things on my own to differentiate myself, but you have to accept that you're going to do those things generally speaking less efficiently than the industry standard because of the economies of scale and the network effects that the infrastructure providers have.

所以你们绕过了这个问题，显然今天你们的用例已经扩展了。我们看到了生成式 AI 的样子，像 OpenAI、Anthropic 等公司。那么，你们在基于人类反馈的强化学习（RLHF）范式中扮演什么角色？你们能把之前用过的类似基本方法应用到现在的生成式 AI 世界吗？

So you got going around that and obviously your use cases have been expanded today. We've seen what generative AI looks like, companies like OpenAI and Anthropic and many others. So, where do you all play in the reinforcement learning, human feedback paradigm? Like can you apply the similar primitives that you got going on there to now this world of generative AI?

是的。我认为现代 AI 最疯狂的一点是，这些模型的大部分能力都是由数据教出来的。你知道，你仍然没有那种只是自己学习、随机展示这些非常人类技能的 AI 系统。它们是通过大规模数据集和人类数据来学习的。所以我们做的是构建我们称之为数据引擎的东西，这基本上是类似的框架，是生态系统中原始数据的炼油厂，但这个数据引擎实际上为当今行业中的每一个领先大语言模型提供动力。基本上每个大语言模型都使用 Scale 的数据引擎，而具体的技术或方法就是你刚才提到的：基于人类反馈的强化学习（RLHF）。

Yeah. So I think one of the craziest things about modern-day AI is that most of the capabilities of these models are taught by data. You know, you still don't have AI systems that are just sort of learning on their own and just randomly demonstrating these very human skills. They're taught to them by large scale data sets and human data. So what we do is we build what we call a data engine, which is basically similar framing, the refinery for raw data in the ecosystem, but that data engine powers every leading LLM in the industry today effectively. Basically every large language model is powered using Scale's data engine and the specific technique or the specific approach is what you just mentioned: reinforcement learning with human feedback.

你能为可能不了解这个术语的人解释一下吗？

Can you explain that for people that maybe don't know that term?

是的，这是一种我们实际上在 2019 年与 OpenAI 合作进行首次实验的技术，但基本方法是：你教模型什么是好的。所以你教模型如何评估一个答案或一个回应是否比另一个更好。它通过大量人类专家教导的例子来学习。所以人类专家会说这个比那个好，原因是什么，模型可以从中学习，然后知道什么是好的。因此，当它真正产生结果时，它有一个内在的关于什么是好、什么是坏、什么比另一个更好的感觉，这被称为奖励模型。然后它进行所谓的强化学习。它基本上利用这种内在的关于什么是好的感觉来优化自己的回应。这意味着这允许模型在很多情况下实际上超越人类表现，因为这有点像世界上每个人都可以是电影评论家，但几乎没有人能制作电影。所以我们每个人都能说出电影可以改进的地方。但显然我不能制作电影。同样，如果人类可以教模型什么是更好的以及如何改进，那么模型就可以不断改进，甚至远远超出人类的能力。

Yeah, so this was a technique that we actually worked with OpenAI back in 2019 on the very first experiments of, but the basic approach is that you teach a model what good looks like. So you teach a model how to assess whether or not one answer or one response is better than another. And it learns that through a bunch of examples where human experts are sort of teaching it. So a human expert will say this one's better than that one and here's why, and the model can learn off of that and then know what good looks like. And so then when it gets to actually producing results, it has an internal sense of what good looks like and what bad looks like and what's better than another thing, what's called a reward model. And then it does what's called reinforcement learning. It basically uses that internal sense of what good looks like to optimize its own responses. And what that means is that this allows the models to actually exceed human performance in a lot of cases because it's kind of like how every human in the world can be a movie critic but almost none of us can make a movie. So each of us can say ways in which a movie could be better or could be improved. But obviously I can't make a movie. In the same way, if humans can teach the model what better looks like and how to improve, then the model can keep improving even far beyond what human capability is.

你处于一个独特的位置，可以看到客户如何利用 AI。在过去几个月或一年里，你有没有什么有趣的轶事或观察，关于企业、大公司同时利用你们和某个模型提供商来做一些你可以说的事情？

You're in such a unique position to see how customers are leveraging AI. Are there any interesting anecdotes or observations you've had in the last couple months or year, whatever it's been about enterprises, big companies leveraging both you and one of the model providers as well to do something that you can speak to?

现在企业面临的一个真正有趣的机遇是，如果你看看当今最先进的模型，它们主要基于公开数据训练，也就是来自开放互联网的数据。但如果你考虑所有可用的数据，99.9% 实际上是某种形式的私有专有数据。一种衡量方式是：我们每个人输入的词汇中，有多大比例最终出现在开放互联网上？微乎其微。大部分都在消息、电子邮件、备忘录里——这些东西永远不会出现在公共互联网上，除非你被传唤。所以大多数企业，无论他们是否意识到，都坐拥远超其他格式或公共互联网上可用数据量的数据宝库。企业的机遇在于，找到方法利用基于公共互联网训练的优质基础模型，然后将它们与自己的数据、业务、客户等上下文融合、微调并专业化，从而产出独特、专有且因过去积累的数据而具有差异化的成果。从广义上讲，这就是世界的发展方向：企业将能够基于专有数据构建具有独特能力的模型。过去几个月令人兴奋的是我们与 OpenAI 等公司的合作，我们也与 Meta 在 Llama 2 上合作。我们构建了一个平台 EGP，使企业能够将自己的数据在 GPT-3.5 或 Llama 2 等基础模型上进行微调，并构建针对自身用例具有独特能力的应用——无论是客户服务、法律应用还是开发。我认为这非常令人兴奋，因为它让企业两全其美：既利用了基础模型提供商的所有惊人发展，又增加了使其独一无二的东西。这是企业未来的范式。虽然还有很长的路要走，但这显然就是未来。

The really interesting opportunity for enterprises now is that if you look at the best-in-class models built today, they're trained predominantly on public data from the open internet. But if you think about the total addressable data, 99.9% of that is actually private proprietary data. One way to benchmark this is: of the words each of us types, what percentage ends up on the open internet? A vanishingly small percentage. Most of it is in messages, emails, memos—things that will never end up on the public internet unless you're subpoenaed. So most enterprises, whether they know it or not, are sitting on troves of data that far exceed the amount accessible in other formats or on the public internet. The opportunity for enterprises is figuring out ways to take great base models trained on the public internet, then intermingle them, fine-tune them, and specialize them on top of their own data, their own business, their own customers, to produce things that are uniquely theirs, proprietary, and differentiated because of all the data they've amassed. Broadly speaking, that's the direction the world is going: enterprises will be able to build models on proprietary data that have unique capabilities. The exciting thing over the past few months is our work with OpenAI and others, and we partnered with Meta on Llama 2 as well. We've built a platform, EGP, that enables enterprises to take their own data, fine-tune it on top of GPT-3.5 or Llama 2 or other base models, and build things uniquely capable for their own use cases—whether for customer care, legal applications, or development. I think this is incredibly exciting because it's a way for enterprises to get the best of both worlds: leveraging all the incredible development from foundation model providers while adding something that makes it uniquely theirs. This is the paradigm of the future for enterprise. There's a long way to get there, but this is clearly what the future will be.

假设你是一家初创公司或企业的高管或创始人，在核心业务中担任决策角色，但与人工智能无关。你现在应该做什么？对于那些不在财富 100 强公司、没有专门人员思考这个问题的人，你有什么建议？普通高管或创始人如何发现他们可能利用人工智能做什么？

Let's say you're an executive or a founder at a startup or an enterprise in a decision-making role in the core business not related to AI. What should you be doing right now? What recommendations would you have for someone who isn't at a Fortune 100 company with people specialized in thinking about it? How does the average executive or founder discover what they could potentially use AI for?

首先，梳理并盘点你独特的数据资产。思考一个思维模型：如果有一个超人，能够比任何人都更快地阅读所有信息，那么这个人能比世界上其他人做得更好的是什么？这是对模型能力的一个粗略近似。模型在存储信息方面比人脑更擅长，而且不受时间限制，所以它们可以阅读所有内容。然后考虑从这样一个系统中你能获得哪些独特能力。进行这个思维练习，思考你能从中做什么独特的事情——既包括降低成本（客户服务是一个明显的例子），也包括进攻性的事情。然后与了解 AI 的合作伙伴（比如我们、OpenAI、Anthropic——这些看到整个生态系统发展的公司）一起构建这些能力。要赶紧去做，因为我相信在很短的时间内，哪些企业拥抱了 AI、哪些没有，就会变得非常明显，从消费者体验和财务数据上都能看出来。

First, go through and catalog what your unique data assets are. Think through a mental model: if there were a superhuman person who could read through all that information more quickly than anyone else, what things would that person be able to do better than anyone else in the world? That's a rough approximation of what the models look like. Models are better at storing information than human brains and are not as time-limited, so they can read through everything. Then consider the unique capabilities you get from a system that has done that. Go through that mental exercise and think about what unique things you can do from there—both cost reduction (customer care is a clear example) and offensive things. Then seek to build those out with knowledgeable AI partners like ourselves, OpenAI, Anthropic—companies that see the entire ecosystem play out. Race to do that because I believe that in a pretty short time frame, it will become very clear which businesses have embraced AI and which haven't, evident from consumer experience as well as financials.

你如何将过去五年人工智能的出现与个人电脑、互联网、智能手机、iPhone 等过去趋势进行比较？你如何看待社会影响、GDP 提升、生产力增益？

How do you compare the advent of AI over the last 5 years to past trends like the personal computer, internet, smartphone, iPhone? How do you think about societal impact, GDP lift, productivity gains?

我诚实的看法是，它将比所有这些都更大。至少，人工智能显然是一种新的消费者范式，是人们期望与技术互动的新方式。从这个意义上说，它至少是另一个移动时代——移动和个人计算是范式和可及性的巨大变革。同样的事情正在发生，AI 聊天机器人作为一种极其流行的技术交付方式。作为基线，你可以将其视为技术的新消费者范式。但关键是，人工智能长期以来一直是一项被大肆炒作的技术，这是有充分理由的——它是解锁人类生产力的圣杯。

My honest take is that it will be bigger than all of them. At minimum, AI is clearly a new consumer paradigm and a new way people will expect to interact with technology. So in that sense, it's at least another mobile—mobile and personal computing were massive changes in paradigm and accessibility. The same is happening with AI chatbots as an extremely popular delivery method for technology. As a baseline, you can view it as a new consumer paradigm for technology. But the upshot is that AI has been a very hyped technology for a long time for good reason—it is the holy grail of unlocking human productivity.

如果你从生产力的角度来考虑，大致就是人均 GDP，经济产出除以人口数量。突然之间，如果有了技术，也就是算法或 AI 系统，能够开始完成原本需要人类完成的相当大一部分工作，那么你就可能获得巨大的生产力提升。另一种看待方式是，美国 GDP 大约是 27 万亿美元，其中软件和 IT 服务约占 2 万亿美元。所以你我花所有时间思考的都是那 2 万亿美元的部分，这虽然不少，但还不到 10%。美国 GDP 中有 16 万亿美元，超过一半，来自服务业，其中最大的两块是医疗和金融服务。这 16 万亿美元服务业 GDP 的潜在颠覆，就是人工智能的潜力所在。在那里，你可以实现 10 倍的生产力提升，10 倍更好的消费者体验，10 倍的经济效率。你无法想象比这更大的经济机遇。在很多方面，我认为这是最大的经济浪潮，直到未来出现某种同样有影响力的新技术。关键问题是：要实现这一点，你只需要相信模型会继续快速进步，因为无论如何，如果模型以现在的速度持续改进，我们最终会进入那个经济颠覆机会前所未有的世界。我认为我们 AI 社区看不到这种放缓很快发生。所以我们正处于世界上有史以来最伟大的经济引擎之一被发明的过程中，这将是我们在技术变革中看到的最特别的事情之一。

If you take the framing of productivity, which is roughly speaking GDP per capita, economic output divided by the number of human heads, all of a sudden if you have technology, algorithms or AI systems that can start doing pretty meaningful chunks of what would otherwise require humans, you have a potentially ridiculous unlock on productivity. Another way to look at this is if you take all of US GDP, it's roughly $27 trillion. Software and IT services is about $2 trillion of that. So everything that you or I spend all of our time on is thinking about that $2 trillion bucket, which is not nothing, but it's not even 10%. $16 trillion of US GDP, more than half, is in services, the biggest buckets of which are healthcare and financial services. The potential disruption of this $16 trillion of services GDP is the potential of artificial intelligence. That's where you can potentially transform to be 10x more productive, 10x better for the consumer, 10x more economically efficient in every way. You can't imagine an economic opportunity bigger than that. In many ways, I think it is the biggest economic wave until obviously the future has some new technology with the ability to be as impactful. The key question is: to unlock that, you just need to believe that the models will keep getting better pretty quickly, because no matter what, if the models keep improving at the rate they're improving now, we're going to end up in that world where the opportunity to disrupt the economy is totally unprecedented. I think we as an AI community don't see that slowdown happening anytime soon. So we're in the midst of potentially one of the greatest economic engines of the world being invented, and that will be one of the most special technological changes we see.

你曾说未来两到三年的 AI 将定义世界未来二三十年。你是什么意思？这与生产力提升有关吗？还是地缘政治方面的？

You said the next two to three years of AI are going to define the coming two to three decades of the world. What did you mean by that? Was that related to a lot of this productivity gain stuff? Was that geopolitical in that comment?

我通常认为，从国家间的力量平衡来看，有两种看待世界的方式。你可以从经济角度和硬实力角度来看。二战前的大部分世界历史是由硬实力决定的，而过去大约 80 年的历史则是由经济实力决定的。你当然可以问未来 80 年将由什么决定，但至少是两者之一。所以在这个框架下，未来两到三年的 AI 发展相当惊人。过去三四年 AI 发展的所有成就都令人震惊。2019 年，GPT-2 数不到 10，会输出胡言乱语，完全无法理解。四年后，GPT-4 可能比世界上大多数人更有说服力和口才。这发生在四年内，模型规模大约扩大了 1000 倍。GPT-2 大约有 20 亿参数，GPT-4 根据估计在 1 万亿到 2 万亿参数之间。我们看到了从蠕虫级别的智能到相当令人信服的人类智能的转变。未来两到三年，许多公司公开表示将进行另一次 100 倍的规模扩张。所以这些公司将从在这些模型上花费数亿美元到花费数百亿美元。我的预期是，这将带来非常强大的算法，能够影响经济实力和硬实力。假设我们处于这种起飞场景。经济实力方面，理由很明确。如果你相信我刚才说的，AI 是全球生产力最重要的东西，那么谁先到达那里，谁最快将其融入经济，谁先利用 AI，哪个国家或社会先做到，就会在经济上获得显著优势。从硬实力角度看，如果你相信这项技术与原子弹类似，我们可以深入探讨，但如果你相信它是那种能够威慑和投射硬实力的技术，那么它也将从根本上改变军事力量平衡。所以我觉得无论你怎么看，这项技术，尽管今天我们把它看作聊天机器人，却是未来 50 年全球力量平衡的核心。

I generally take the stance that there are two ways to look at the world in terms of the balance of power between countries. You can look at it from an economic standpoint and from a hard power standpoint. Probably most of the history of the world before World War II was dictated by hard power, and then most of the history for the past 80 or so years has been dictated by economic power. You could certainly ask which will define the next 80 years, but at minimum it's one of the two. So with that framing, one of the things that's quite shocking is the next two to three years of AI development. Everything we've seen over the past three or four years of AI development is shocking. In 2019, GPT-2 couldn't count to 10, it would spit out gibberish English, it was totally unintelligible. Then four years later, GPT-4 is probably more convincing and eloquent than most people in the world. That happened over the course of four years and roughly a thousand-fold scale-up of the models. GPT-2 is roughly 2 billion parameters. GPT-4, depending on estimates, is somewhere between 1 to 2 trillion parameters. It's been a roughly thousand-fold scale-up. We've seen this transformation from worm-level intelligence to something quite convincingly human. The next two to three years, many companies are on record for undergoing another 100x scale-up. So these people will go from spending hundreds of millions of dollars on these models to tens of billions of dollars. My expectation is that that's going to deliver very powerful algorithms that have the ability to impact both economic power and hard power. So let's say we're in this takeoff scenario. Economic power, the case is pretty clear. If you believe what I just said around it being the most important thing for global productivity, then whoever gets there first, whoever integrates into their economy the fastest, whoever is able to leverage AI first, whichever country or society does that first is going to have a meaningful leg up from an economic standpoint. From a hard power perspective, if you believe the technology is of a similar vein as the atomic bomb, which we can certainly dive into, but if you believe it's that kind of technology with the ability to both deter and project hard power to that degree, then it's also going to fundamentally change the balance of military power. So it feels to me like no matter how you slice it, this technology, while today we think about as a chatbot, is at the core of the balance of power globally for the next 50 years.

你同意原子弹类比中的哪一部分？拒绝哪一部分？你显然对此有熟悉感，因为你在洛斯阿拉莫斯长大。你相信这个比较的哪些方面，哪些不相信？

What part of the atomic bomb analogy do you agree with? What part do you reject? You obviously have familiarity with elements of it, growing up in Los Alamos. What do you believe about that comparison versus not?

这里有很多有趣的细微差别。原子弹显然主要是一种战争武器。它是一种武器，全世界很快就能达成共识，我们不想再使用它了。所以它很快从几次使用变成了冲突的明确威慑，成为全球的巨大稳定器。人工智能的不同之处在于，无论如何，我们都必须将这项技术用于经济目的。所以不存在世界各国聚在一起说‘嘿，我们不再使用 AI 了’的场景。而且人工智能是一项很难检测其使用的技术。

There are a bunch of interesting nuances here. The atomic bomb was obviously primarily a weapon of war. It is a weapon, and it's something that pretty clearly we as an entire world could pretty quickly agree we didn't want to use anymore. So it very quickly went from a few uses to being this very clear deterrent for conflict, and this huge stabilizer for the globe. The difference with artificial intelligence is that no matter what, we're going to have to use the technology for economic purposes. So there's no scenario in which the countries of the world are going to get together and say, 'Hey, we're not going to use AI anymore.' And artificial intelligence is a pretty difficult technology to detect the use of.

所以，AI 的问题部分在于，俄罗斯今天可能就在用它进行网络攻击，而我们很难真正知道他们在这么做。但使用核弹几乎不可能隐藏，对吧？正因如此，围绕这项技术的使用、公平使用以及如何制定恐怖分子使用技术的标准，就变得非常困难。所以我认为这给世界带来了挑战：AI 是一种很难被限制在某种盒子里的技术，不像核武器。我认为相似之处在于，这是一项技术曲线非常陡峭、规模化收益非常明显的技术。因此，如果美国能够成为领导者，或者一小部分民主国家能够成为这项技术的领导者，那么我认为它确实有潜力对那些落后于曲线的国家形成巨大的威慑。这在原子武器和核武器上已经得到了验证。

So, part of the issue with AI is that Russia could be using it for cyber attacks today and it'd be very hard for us to actually know that that's what they were doing. It's almost impossible to hide the fact that you used a nuke, right? So because of that, it makes it pretty hard to set the right international standards around the use of the technology, the fair use of the technology, how you set standards around how terrorists can and should use the technology. And so I think that presents challenges as a world, which is that this is a hard technology to keep in any sort of box, unlike nukes. The ways in which it's similar I think are that it's a technology that has a very steep technological curve and has very clear benefits to scale. So to the degree that the United States can be the leader, or a small set of democratic countries can be the leaders in this technology, then I do think it has the potential to be a huge deterrent towards other countries that are behind on that curve. And that's certainly been the case with atomic weapons and nuclear weapons.

你是否担心那种快速起飞带来的灾难性风险，以及 AI 本身更邪恶的一面，而不是被外国对手用于生物武器之类的东西？你是否担心 AI 本身？

Do you worry about the catastrophic risk scenario of that fast takeoff and anything more nefarious with AI itself, rather than being used by foreign adversaries for things like bioweapons or whatever you want to compare it to? Do you worry about it in and of itself?

我对 AI 风险的分类大致有三个桶。第一个桶是 AI 本身的风险，即 AI 本身成为人类的威胁。就我个人而言，这不是我最担心或最关注的桶。我可以多说一点。第二个是 AI 滥用类别，即威权国家或恐怖组织滥用这项技术。我认为这是一个非常真实的风险。我认为这是我们面临的最真实的风险。最后一个风险是二阶效应：大规模劳动力转移会导致各种政治不稳定、国内不稳定、民粹主义、社会趋势在许多发达国家出现。所以滥用风险我认为非常真实。我认为全球恐怖主义总体在上升，而技术被滥用的可能性非常高，再次针对网络攻击、生物攻击、生物武器、信息战，甚至像这样的版本我认为几乎是最直接或最清晰的：有像 Character AI 和 Replica 这样的公司，你可以拥有一个 AI 模型，成为各国很大比例公民的真正伴侣。如果你有一家外国运营的 AI 伴侣公司，我认为那可能是你能拥有的最有效的情报机构。所以在 AI 滥用领域有很多值得担忧的地方。我认为这当然非常令人担忧。这是我们作为一个国家、一个社会需要思考的问题。我们如何减轻这些风险？拜登政府曾有一项行政命令。我认为我们肯定在考虑这些。

My taxonomy on AI risks is sort of like there's three buckets. The first bucket is the AI qua AI risk. So the AI itself becomes the threat to humanity. That's personally speaking not the bucket that I am most worried about or concerned about. And I can speak more about that. There's the AI misuse category. So authoritarian countries or terrorist groups misusing the technology. I think that's a very real risk. I think that's the most real risk that we have. And then there's the last risk, which is a second-order effect: with massive labor displacement, you'll see all sorts of political instability, domestic instability, populism, social trends in many developed countries. So the misuse one I think is very real. I think we're seeing overall an increase in terrorism in the globe, and I think the potential for misuse of the technology is very high again for cyber attacks, for bio attacks, bioweaponry, for information warfare, and even stuff like the version of this I think is almost the most direct or clear is like there's these companies like Character AI and Replica where you can have an AI model that becomes a genuine companion to huge percentages of the citizens of various countries. If you had a foreign-run and foreign-operated AI companion company, I think that's the most effective intelligence agency that you could possibly have. So there's a lot to be worried about in the realm of AI misuse. And it's something that I think is certainly very concerning. It's something that we as a country, we as a society need to think about. How do we mitigate those risks? There was the executive order from the Biden administration. I think we're certainly thinking about those.

你认为未来五年人工智能有哪些不可避免的事情，可能不是主流观点，或者普通人不会完全理解？

What's something that you believe inevitable about artificial intelligence in the next 5 years that maybe isn't mainstream or the average person wouldn't fully appreciate?

我想提几件事。第一件，AI 领域的大多数人都看到并相信，但肯定还没有完全成为主流的是，这些模型将很快成为大多数国家最大的投资之一。所以，如果你相信这些投资从数亿美元到数十亿美元再到数百亿美元，我的意思是，没有多少国家能够负担得起千亿美元的投资，无论是通过私营行业还是公共部门、政府本身资助。所以这很快就成为世界上最大的经济项目或科学项目之一，我认为人们可能惊讶于它还没有达到那个程度。你知道这些模型已经花费了数亿美元，但很多人能负担得起几亿美元。很快，在投资规模上，它将几乎像粒子加速器或这些大型科学项目。我认为另一件很多人没有考虑或只是慢慢融入的事情是，人类与他人互动的时间与直接与模型互动的时间的比例，这个比例将不断加速向模型倾斜。所以，除了监管之外，真的没有理由相信未来几十年里我与模型互动的时间比例会在任何一点下降。所以它会单调递增。对我来说已经很高了。我已经经常与 ChatGPT 互动。我认为这是一个非常奇怪的社会学场景需要我们应对，那就是无论如何，这些模型将开始侵占你与他人交谈和互动的时间。如果你相信 AI 模型只会变得更好，如果你相信它们只会拥有更有趣的数据，如果你相信产品会变得更好，那么改进的单调性将是非常奇怪的。也许这些不会在未来五年发生，也许会在十年或十五年内发生。谁知道它们什么时候发生，但在某个时候，人们将花费超过一半的时间与模型交谈而不是与人类。

I think there's a bunch of things I'll mention. I think one that most people in AI see and believe, but certainly is not fully mainstream, is just that these models are going to become very quickly some of the largest investments in most countries. So, if you believe these go from hundreds of millions of dollars to billions of dollars to tens of billions of dollars to hundreds of billions of dollars. I mean, there's not that many countries that can afford a hundred billion dollar investment either funded through private industry or funded through the public sector, through the government itself. And so this very quickly becomes one of the largest economic projects or scientific projects that the world has seen, which I think is maybe surprising to people that it isn't that yet. You know these models have cost hundreds of millions of dollars, but a lot of people can afford a few hundred million dollars. Very quickly it's going to be almost like particle accelerators or these massive scientific projects in terms of scale of investment. I think the other piece that many people don't think about or is just going to slowly blend in is that the percent of time that humans are interacting with other people versus to a model directly, that split is just going to keep accelerating in the direction of the model. So, there's truly no reason outside of regulation you would believe that the percent of time I spent interacting with models is going to decrease at any point for the next few decades. So that's going to increase monotonically. It's already pretty high for me. I interact with ChatGPT quite a bit already. And I think that's a very weird sociological scenario for us to contend with, which is no matter what, these models are going to start eating into all the time you spend talking and interacting with other people. If you believe that the AI models are only going to get better, if you believe that they're only going to have more interesting data, if you believe the products are going to get better, that the monotonicity of the improvement is going to be very weird to think about. And maybe these don't happen in the next five years, maybe these happen over 10 years or 15 years. Who knows when they happen, but at some point people are gonna spend more than half their time talking to models versus humans.

你怎么看这种正交性，即 AI 擅长什么、不擅长什么，而且它偏离了正轨？

How do you think about that orthogonality and what AI is good at versus what it isn't and it's off to the side?

是的，我认为这完全归结于数据可用性。回顾一下，数据是所有算法的生命线。它们学到的一切、能做的一切，都是从数据中学来的。结果，通过过去几十年使用互联网、在 Reddit 上评论、上传东西到网上，我们碰巧创造了有史以来最大的人类行为数据集。所以，我们在电脑上做的任何事情——大部分是知识工作、知识相关或智力活动，因为本质上它们是从现实世界抽象出来的——模型都有大量数据。但它们在诸如拿起东西、扔球、制造东西等具身于现实世界的事情上，数据非常少。这种情况会持续很长时间。这些模型的数字存在和数字智能，很可能永远都会超越物理具身能力。从数据可用性的角度看，这完全说得通。而真正奇怪的地方在于其经济影响，以及这对劳动力未来意味着什么。

Yeah, I think this just all boils down to data availability. So, going back to it, right? Data is the lifeblood of all these algorithms. Everything they learn, everything they are capable of, they've learned from data. And so it turns out that by using the internet over the past few decades, and by commenting on Reddit and uploading stuff to the internet, we've happened to have been creating the largest dataset of human behavior ever. So anything that we did on a computer, which most of it was fundamentally knowledge work or knowledge related or intellectual, because by definition it's abstracted away from the real world, that's what the models have a lot of data on. So they have remarkably little data of what it's like to pick something up or throw a ball or manufacture something, or all the things that are embodied in the real world. There's very little data on that, and that's going to be true for a long time. The digital presence of these models and digital intelligence is always going to surpass, probably for perpetuity, the physical embodied capability. If you think about it from a data availability standpoint, I think it makes perfect sense. And obviously where it gets really weird is the economic impacts of this and what that means for the future of labor.

你提到了模型开发的三个组成部分，我想是人才、算力和数据。你认为今天最大的限制因素是什么，5 年或 10 年后又会是什么？

You touched on the three components of model development, I guess, being talent, compute, and data. What do you think the most limiting factor is today, and what do you think it will be in 5 years time or 10 years time?

我认为数据和算力绝对是今天的限制因素。算力由于制造能力有非常明确的限制。两者的供应链都值得深入探讨。目前，驱动这些模型的高端 GPU 100% 在台湾制造。台积电投入了数百甚至数千亿美元的资本支出建造这些晶圆厂，并不断改进。这对这些模型的算力能力和容量构成了非常强的上限。因此，如果你相信持续的指数级 Scaling，除非供应链也指数级扩张，否则很难实现，而这在今天并不现实。所以算力既是今天的瓶颈——显然我们看到英伟达芯片的售价以及初创公司对它们的渴望——也是指数增长场景的明确限制因素。数据也是如此。很多人观察到是否还有更多的预训练数据，我们是否已经用完了高质量 token？有一些非常清晰的论点表明，缩放定律将难以维持，因为互联网上已经没有那么多高质量数据了。还有一个争论：视频数据是否高质量？文本是一种异常压缩的知识和信息形式，视频则压缩程度低得多。所以如果没有足够的预训练数据，很多缺口是通过 RLHF 和后训练数据来弥补的。我认为我们将开始看到类似的瓶颈：真正需要的人类专家数量来推动 RLHF 阶段——人类专家本身就成了 GPU——这些推动模型改进的人类专家的数量和质量将成为行业的另一个供应链瓶颈。

I think data and compute are definitely the limiting factors today. Compute has a very clear limit because of manufacturing capability. So the supply chains for both are worth diving into. 100% of high-end GPUs that fuel these models are manufactured in Taiwan today. There are these fabs that TSMC has put tens if not hundreds of billions of dollars into capex to build and continue to refine and improve. That's a very strong upper bound limit for the compute capability and capacity for these models. So by definition, if you believe in continued exponential scaling, it gets pretty hard unless you have an exponential scaling of the supply chain as well, which is not really feasible today. So compute is both the pinch point today—obviously we see how much Nvidia chips sell for and how much startups want them—but also a clear limiting factor to the exponential growth scenario. Data is as well. A lot of people observed whether there is more pre-training data out there, have we run out of high-quality tokens? There are some very lucid arguments that show that some of the scaling laws will be tough to keep up because we just don't have that much more high-quality data on the internet. There's this argument: is video data high quality or not? Text is a very unusually compressed form of knowledge and information. Video is much less compressed. So if you don't have enough pre-training data, a lot of this has been made up for in RLHF and post-training data. I think we're going to start seeing similar kinds of bottlenecks where the amount of human experts who are really needed to fuel the RLHF stages—human experts become GPUs in their own right—the number and quality of human experts fueling model improvement is going to become another supply chain bottleneck for the industry.

从 GPT-2 到 3 再到 4，从外部看几乎像是线性发展，但显然这些更像是阶梯函数。你认为在现有约束下，我们是否会在某个点遇到平台期，需要其中一个因素有更大的突破才能达到下一个主要台阶？

As we've looked at GPT-2 to 3 to 4, it seems to the outside almost like linear development, but clearly these are more stairstep functions along the way. Do you think with the constraints we have, we're going to hit some plateau at some point that's going to require a much bigger unlock of one of these things to really reach that next major step function?

当你与领先实验室的人交谈时，他们把所有时间都花在思考这些模型的供应链上。所以我认为，隐含地，如果什么都不做，这些将是巨大的瓶颈。但话虽如此，我认为这可能是我们见过的最伟大的人类工程项目。所以我认为我们会找到解决办法。这意味着你会开始看到一些相当疯狂的行动，试图确保供应链能够继续 Scaling。但同样，我认为这是我们所处的技术必然性。

When you talk to people at the leading labs, they spend all their time thinking about the supply chains for these models. So I think that implicitly, if nothing happens, these will be really big bottlenecks. But that being said, I think that this is potentially the greatest human engineering project we've ever seen. So I think we're going to figure things out. I think what that means is you're going to start seeing some pretty crazy actions to try to secure and ensure that the supply chains can continue scaling. But again, I think that's the technological imperative we operate in.

你认为我们作为社会是否低估了对台湾的依赖以及台湾所处的政治地位，这对人工智能意味着什么？

Do you think we're underappreciating as a society the reliance on Taiwan and the political position that Taiwanese find themselves in, and what that means for artificial intelligence?

一个非常明显的迹象表明我们低估了这一点，就是英伟达和台积电之间的估值倍数差距。台积电的估值倍数远低于英伟达。英伟达当然是利润率更高的公司，所以部分差距是合理的，但据我与公开市场投资者的交流，台积电因为地缘政治风险而受到打压。台湾正处于世界的这个压力点上。

One very clear indication of the degree to which we don't appreciate it is just in the multiple gap between Nvidia and TSMC. TSMC trades at a dramatically lower multiple than Nvidia. Nvidia is a higher margin company, of course, so some of it is well-deserved, but TSMC, from my talking with public market investors, gets dinged because of this geopolitical risk. Taiwan is just at this pressure point for the world.

你对开源与闭源模型有什么看法？这似乎是当前的一个大辩论。你作为公司有什么意见吗？

What's your perspective on open source versus closed source models? It seems to be a big debate these days. Do you have any opinions on that as a company?

我个人观点是对技术如何发展持相当不可知的态度。我认为 AI 是一项极其强大且有益的技术。我认为这些模型的所有发展都很好，只要你有安全的开源开发和安全的闭源开发。两者都可能做得不好且不安全，也都可能做得安全且出色。如果两者都安全发展，那就很好。

My personal point of view is to be quite agnostic to how the technology develops. I think that AI is an incredibly powerful and good technology. I think that all development on these models is great, as long as you have safe open source development as well as safe closed source development. Both can be done poorly and unsafely, and both can be done safely and well. If you have safe development on both, it's great.

我认为开源模型可能是确保 AI 实现其全部经济影响的必要条件。在很多场景下，你需要一个在某个地方运行的小模型，这很可能需要某种形式的开源模型。为了满足这种需求而搞一个小型闭源模型是没有意义的。所以我认为，拥有开源模型对经济增长和经济繁荣是有好处的。

I think that open source models are probably a requirement to ensure that AI achieves the full economic impact that it can have. There's a lot of scenarios where you need a small model running somewhere, and that probably needs to be an open source model of some form. It doesn't make sense for there to be some small closed source model just to fit that need. So I think it's good for economic growth and economic prosperity that we have open source models.

我听过你谈论 AI 的竞争曲线。你能多谈谈不平等以及由于缩放定律导致的规模扩张和民主化之间的竞争曲线吗？

I've heard you talk about the competing curves of AI. Can you talk about inequality and the competing curves of scale and democratization a little bit more because of the scaling laws?

随着模型训练成本变得极其昂贵——数百亿、数千亿美元，未来甚至可能达到数万亿美元——这明显限制了底层技术的可及性，就像我们谁都无法使用粒子加速器一样。那是非民主化技术的典型代表。所以这是技术发展的一个主要支柱。另一个支柱是如何如此迅速地推动这些模型沿着成本曲线下降，以至于在拥有极其强大的闭源模型几年后，你就有了非常好的开源模型，它们沿着成本曲线急剧下降。我认为我们在开源模型中看到了这一点。例如，GPT-3.5 级别的模型出现得非常快，而且实际上非常小。最近的一些研究表明，100 亿参数甚至更小的模型可以达到 GPT-3.5 的水平。所以这种民主化的快速改进意味着你有一条规模扩张曲线，另一条是从前沿模型结果到民主化的速度曲线。这些是整个行业的推拉力量。

As the models become ridiculously expensive to train—tens of billions, hundreds of billions of dollars, potentially even trillions in the future—that very clearly limits the accessibility to the underlying technology, just like none of us have access to particle accelerators. That is the poster child of non-democratized technology. So that's one major tent pole for how the technology develops. The other one is how do you push all these models down the cost curve so quickly that a few years after you have incredibly powerful closed source models, you have very good open source models that climb down the cost curve dramatically quickly. I think we're seeing that in open source models. For example, GPT-3.5 level models have happened very quickly and are actually very small. Some recent things show that 10 billion parameter or even smaller models can perform at the level of GPT-3.5. So this rapid improvement of democratization means you have one curve of scaling and another curve of speed from frontier model result to democratization. These are the push and pull of the entire industry.

什么是图灵陷阱？在你看来，它为什么重要？

What is the Turing trap and why is that significant in your mind?

图灵陷阱来自经济学家埃里克·布林约尔松等人撰写的一篇优秀论文。基本前提是，AI 的起始条件来自图灵测试的概念——在什么情况下 AI 能够完全模仿人类。由于这种框架，我们主要将 AI 视为人类的替代品。所以我们认为 AI 会取代劳动力中的人类，而这将是它对经济的影响。但正如布林约尔松教授所论证的，这是一个陷阱。实际发生的情况是，AI 系统将沿着能力曲线缓慢上升，随着它们的到来，大部分价值将来自人机混合系统。通过人类能力和 AI 能力之间非常有趣、复杂和微妙的互动，你才能获得这些非常有经济价值的东西。正因为如此，在大多数情况下，AI 最终会成为更多就业机会的净创造者，或者是对人类劳动力更多需求的净创造者。这是非常重要的信息之一：有一种看法认为 AI 会抢走我们所有的工作，但答案是 AI 将创造一种根本不同的经济，拥有不同的工作组合，这很可能会净创造对人类劳动力的更大需求。

The Turing trap comes from a great paper by economist Erik Brynjolfsson and others. The basic premise is that the starting condition of AI came from the concept of the Turing test—at what point do you have an AI that can fully imitate a person. Because of that framing, we've thought about AI as a replacement for humans predominantly. So we think about AI replacing humans in the workforce, and that will be its impact on the economy. But as Professor Brynjolfsson argues, that is a trap. What's actually going to happen is that AI systems will slowly walk up the capability curve, and as they come in, most of the value will be generated from hybrid human-AI systems. Through very interesting, complex, and nuanced interactions between human capability and AI capability, you get these very economically valuable things. Because of that, AI in most outcomes ends up being a net creator of more jobs or a net creator of more demand for human labor. That's one of the very important messages: there's this perception that AI will just take all our jobs, but the answer is that AI is going to create a fundamentally different economy with a different mix of jobs, which will probably net create greater demand for human labor.

AI 有很多复杂性和细微差别。还有没有另一个普遍的误解，你想澄清或表达你的不同意见？

There's a lot of complexities and nuances associated with AI. Is there another general misconception that people have that you would like to clarify or express your dissenting opinion on?

人们在直觉上思考 AI 时犯的一个主要错误是，很容易说‘你有一个 AI 系统，你使用 GPT-4，你发现它总是产生幻觉，然后你摊手说这项技术有根本性局限，永远不会成功，因为它会产生幻觉。’AI 的棘手之处在于很难对它下注，因为之前的每一个例子中，当你使用早期版本的模型时——如果你使用 GPT-2 并说‘这东西连数到 10 都不会’，你摊手说这里没有未来；或者 GPT-3，你用它，它连一个简单的数学题都解不了，你摊手说这不会成功。我认为很多人，甚至 AI 行业内部的人，从根本上并不真正相信模型的改进。这很遗憾，因为现实是模型会变得好得多。很难想象它们会如何变得更好，但它们会的，我们需要思考一个我们处于持续模型改进轨道上的世界。

One major thing people get wrong when they think intuitively about AI is that it's very easy to say, 'You have an AI system, you use GPT-4, you realize it just hallucinates all the time, and then you throw your hands up and say this technology is fundamentally limited and never going anywhere because it hallucinates.' The tricky thing about AI is it's very hard to bet against because every prior instance where you used an earlier version of models—if you used GPT-2 and said 'this thing can't count to 10,' you throw your hands up and say there's no future here; or GPT-3, you use it and it can't solve a simple math problem, and you throw your hands up and say this isn't going anywhere. I think a lot of people even in the AI industry fundamentally don't actually believe in model improvement. It's a shame because the reality is the models are going to get a lot better. It's hard to imagine how they get a lot better, but they will, and we need to be thinking about a world where we're on this continued track of model improvement.

你研究地缘政治以及 AI 如何在其中发挥作用。你最近就这个话题做了 TED 演讲。你能谈谈你看到的 AI 在地缘政治世界中展开的斗争，特别是中美之间吗？

You're a student of geopolitics and how AI plays into that. You recently did a TED talk on the subject. Can you speak to the battle you see playing out in AI within the geopolitical world, particularly between China and the US?

AI 令人惊讶的一个方面是，它在多大程度上成为世界上许多国家的明确目标和当务之急。显然，它的大部分是在美国发明的，在谷歌、OpenAI、DeepMind 等公司。但很快，你看到中国试图快速行动。中国科技巨头总共购买了价值超过 50 亿美元的 NVIDIA 芯片。那是很多芯片。你看到阿联酋，特别是阿联酋和沙特阿拉伯，非常积极地进入这项技术，建设大型数据中心。阿联酋已经连续开源了两个开源模型。

One of the ways AI has been surprising is the degree to which it's become a clear objective and imperative for many countries around the world. Obviously, much of it was invented in the United States at Google, OpenAI, DeepMind, etc. But very quickly, you see China trying to move very fast. The Chinese tech giants have bought an aggregate of over five billion dollars worth of NVIDIA chips. That's a lot of chips. You see the UAE, particularly the UAE and Saudi Arabia, moving very aggressively into the technology, building large data centers. The UAE has open-sourced two successive open source models.

其中一个模型有 1800 亿参数。他们正在构建非常庞大且严肃的模型。在欧洲，我们看到一些最好的开源模型来自欧洲公司和初创企业。从我与许多其他国家人士的交流来看，还有很多国家在人工智能方面有着明确的雄心。所以至少，这项技术正被许多国家视为对未来至关重要。更令人担忧的是，某些国家，特别是中国，非常清醒地认识到这项技术可能产生的巨大影响。有许多解放军文件明确提到，人工智能和其他突破性技术可以让解放军超越其对手，尤其是世界上最强大的军事力量——美国。因为我们会在传统平台上过度投资，只是升级它们，而不是投资新的突破性技术。他们会过度投资新的突破性技术，并可能超越我们，就像中国在金融科技和支付技术领域超越美国一样——大多数人认为中国的数字支付基础设施已经超过了美国的支付基础设施。问题是，全球 GDP 的 40%呢？美国和中国加起来占全球 GDP 的 40%。所以它们是经济中的两个庞然大物。关键问题是，人工智能是中国超越美国，或者至少大幅缩小差距的催化剂吗？还是说，这项技术能让美国通过延续和维持“美国治下的和平”来确保全球稳定？如果你和许多政治学家交谈，他们会有一个相当明确的共识：如果中国的军事能力赶上美国，那将是一个非常不稳定的世界。无论你站在哪一边，这都肯定会导致更大程度的全球不稳定。过去 80 年相对和平的一个重要原因，是美国一直是世界上明确的硬实力超级大国。如果有两个超级大国，系统内的熵就会增加。会有更多的代理人战争，更多的不稳定，更多的战争，更多的死亡。所以，在民主与威权主义以及不同的政府体制和世界组织方式之间的这场更广泛的斗争中，人工智能是其中的一个重要棋子。这就是为什么我认为我们作为美国人，或者说美国整体，能够保持领先地位至关重要。也许可以谈谈中国与美国目前支出比例的问题。至少在过去几年里，中国，也就是解放军，一直将其预算的大约 1%到 2%用于人工智能技术。在同一时期，美国国防部将我们预算的 0.1%到 2%用于人工智能技术。所以解放军一直预测的情况现在正在现实中上演：我们在传统平台和传统技术上过度投资，而在突破性技术上投资不足。他们可能会在我们之前取得突破，而我们可能会陷入一种我们不喜欢的局面。

One of which is 180 billion parameters. These are very big and serious models that they're building. In Europe, you're seeing some of the best open source models coming from European companies and startups. From my conversations with people from many other countries, there are many others who have clear aspirations in AI. So at minimum, it's becoming a technology that a lot of countries are looking at as really important for their future. What's more concerning is the degree to which certain countries, particularly China, are very clear-eyed about the monumental impact this technology can have. There are a number of PLA documents that talk explicitly about how AI and other breakthrough technologies could allow the PLA to leapfrog its adversaries, most notably the United States, which is the most powerful military in the world. Because we're going to overinvest in our legacy platforms and just upgrading them, versus the new breakthrough technologies. They'll overinvest in new breakthrough technologies and they could leapfrog us, just like China leapfrogged the United States in fintech and payments technology, where most people believe their digital payment infrastructure surpasses the state of payments infrastructure in the United States. The question is, what about the 40% of global GDP? The US and China together are 40% of global GDP. So these are the two behemoths in the economy. The key question is, is AI the catalyst for China to overtake the United States, or at minimum dramatically gain ground? Or is it the technology that allows the United States to ensure we can maintain global stability by persisting and continuing Pax Americana? If you talk to a lot of political scientists, there's a pretty clear consensus that if Chinese military capabilities catch up to those of the United States, that's a very unstable world. Whichever side you're on, that definitely results in greater levels of global instability. A major portion of the last 80 years of relative peace has been because America has been the clear hard power superpower in the world. If you have two superpowers, you get a higher level of entropy in the system. There are more proxy wars, more overall instability, more war, more death. So in this broader battle between democracy and authoritarianism and different government systems and ways the world can organize, AI is one of the major chess pieces in that game. That's why I think it's critical that we as Americans, or America in general, is able to maintain that pole position. And maybe speak to the proportionality of what China is spending versus the US today. For the past few years at least, China has been spending, the PLA, the Chinese military, has been spending between roughly 1 and 2% of their budget on AI technologies. In that same time period, the US DoD has been spending 0.1 to 2% of our budget on AI technologies. So what the PLA had been forecasting is actually playing out in reality right now, which is we are overinvesting in our legacy platforms and legacy technologies, underinvesting in the breakthroughs. They might reach a breakthrough before us, and we might be left in a situation we don't like.

当人们听到这个的时候，你已经参加完英国人工智能峰会了。顺便说一句，我觉得你在那里做得很好，真的非常出色。你明天就要出发了。为什么参加这个对你很重要？你希望实现什么？

By the time people hear this, you will have already been to the UK AI summit. I think you did a great job there, by the way. I think it was really well done. You're heading out tomorrow. Why is attending this important to you? And what are you hoping to accomplish?

这里有几个我觉得有趣的线索。一个是确保有一条全球合作的人工智能轨道。我认为，无论你相信什么，如果这项技术像我以及许多人认为的那样重要，那么它就需要世界上许多国家进行清晰、开放的对话。你不希望任何国家偏离轨道，以对世界其他地区不透明的方式行事。这肯定会加剧不稳定。所以至少，所有国家之间就这项技术进行开放对话，对世界具有巨大的内在价值。许多国家都会参加，这很好，感谢英国政府创建了这样一个论坛。另一个关键点是确保我们思考的是这项技术的正确风险。我认为我们谈论了很多前沿风险和存在性风险，但我想确保我们也充分考虑了滥用的风险，以及我们对此做了什么、如何思考。所以对我来说，确保我们对特别是地缘政治风险有一个更广阔的视野，并以此塑造围绕这项技术的全球对话，这很重要。

There are a few threads here that I think are interesting. One is ensuring that there is a track for global cooperation on AI. I think regardless of what you believe, if this technology is as important as I think it is, as many think it is, it's something that requires many countries in the world to have a clear and open dialogue around. You don't want anybody going off track and doing things in a way that is opaque to the rest of the world. That's certainly a driver of instability. So at minimum, there's a huge amount of intrinsic value to the world by having an open dialogue between all the countries to discuss the technology. Many of the countries are going to be there, which is great, and kudos to the UK government for creating such a forum. The other piece that's critical is ensuring that we're thinking about the right risks of the technology. I think we talk a lot about the frontier-level risks and some of the existential risks, and I want to make sure that we're also thinking a lot about the risks of misuse and what we are doing about those and how we think about those. So it's important to me to ensure that we have a broader view of particularly the geopolitical risks at play and that shapes the sort of global dialogue around the technology.

你认为政府在监管人工智能方面扮演什么角色？

What role do you think the government plays in regulating AI?

是的，我认为这显然是当下的问题，尤其是行政命令已经出台。到目前为止，做法是对技术采取非常轻度的监管，特别是因为我们处于非常早期的阶段。对于像人工智能这样潜力巨大的技术，最糟糕的事情之一就是过早过度监管而浪费机会。所以我认为这很明智。我认为关键是政府需要确保技术的滥用，或者技术可能对消费者或公民造成重大伤害的方式，不会发生，或者至少这些行为会受到严厉惩罚并受到限制，非常有限且难以实施。为此，我认为行政命令的一个关键部分是确保对人工智能系统有适当的测试和评估机制。

Yeah, I think it's obviously the question of the day, literally with the executive order coming out. So far, the approach has been to take a very light touch on regulation of the technology, particularly because we're in such an early stage. One of the worst things you can do for a technology as high potential as AI is to squander the opportunity early on by overregulating it. So I think that's been smart. I think the key is that the government needs to ensure that the misuses of the technology, or the ways in which the technology can be used to create meaningful consumer harm or meaningful harm to the citizen base, don't happen, or that at least those are very highly punished and limited, very limited and difficult to do in some way. To that end, I think a key part of the executive order is ensuring that there's the proper testing and evaluation regime for AI systems.

那么，我们作为社会，如何就某些 AI 系统、用例和应用是否适合用途、是否准备好投入实际使用，还是完全不合格达成一致？这在各种生态系统中都有类似情况。比如 FDA 批准药物，我们不能随便从网上买些分子然后吃下去，还指望没事。飞机、汽车等潜在危险技术也有类似监管。甚至苹果对 App Store 中的应用也有类似做法，必须经过批准。所以我认为这是关键问题。我和白宫的人交流时，都觉得这是一个需要存在但尚未存在的行业。Scale 正努力在这方面发挥重要作用。几个月前，我们与白宫和 Defcon 合作，对这些模型进行了首批公开评估。我们的观点是，需要有一个相当清晰的体系：私营部门的测试者，公共部门给出的明确监管和指导方针，以及模型提供商和 AI 技术实施者的明确选择加入。

So how do we as a society agree that certain AI systems and use cases and applications are fit for purpose and ready for prime time versus totally inadequate? And there are versions of this that exist in all sorts of ecosystems. So the FDA approves drugs. We can't just buy random molecules off the internet and ingest them and expect that to go well. There's similar kinds of regulation on planes, obviously, and cars, and these technologies that are potentially very dangerous. Even Apple does a version of this for apps in the App Store. You have to be approved by the App Store. So I think this is the key question. In my conversation with folks in the White House, it's like this is the industry that needs to exist that doesn't exist. And we at Scale are trying to play a big part in this topic. We worked with the White House and Defcon on some of the first public evaluations of these models a few months ago. Our view is that you need to have a pretty clear regime of testers in the private sector with pretty clear regulation and guidelines given by the public sector, and a very clear opt-in from the model providers and those implementing AI technology.

我想回到 Scale 的创立，从那些更广泛的话题转过来。那么，这个业务背后的最初洞察是什么？你当时认识到了什么，导致了它的创立？

I want to back up to the founding of Scale and transition from some of those broader topics. So what was the original insight behind the business? What did you recognize at that time that sort of led to its founding?

是的，我认为关键洞察很简单：如果 AI 要发展，对数据的需求将呈指数级增长。我当时不知道这会在什么时间框架内发生，或者何时，或者以何种规模和程度，但我非常确信神经网络和 AI 将变得越来越普遍。如果你相信这一点，你就会相信必须有数据基础设施来应对这一挑战和增长。这确实发生了，甚至以令我们惊讶的方式：这些 AI 系统所需的数据量和对新数据的渴望，远远超出了我最初认为在这个时间框架内可能的程度。

Yeah, I think the key insight was that, simply put, if AI were going to grow, the needs on data were going to grow exponentially. I had no idea at what time frame that was going to happen, or when, or at what scale and magnitude, but I had pretty strong conviction that neural networks and AI were going to be more and more ubiquitous. If you believe that, you believe there had to be infrastructure for data to meet that challenge and meet that growth. That certainly played out, I think even in a way that's been surprising to us, which is that the amount of data required for these AI systems and the hunger for new data has far exceeded what I originally would have conceived possible by this time frame.

那么你在 Quora 待了多久才去上学？

And so you spent how long at Quora before going to school?

我在新墨西哥州洛斯阿拉莫斯长大。父母是实验室的物理学家。那里有很多物理学家。然后我去 Quora 工作了大约一年。那是我对技术世界的初次尝试和体验。我 17 岁在那里工作。这非常令人大开眼界，就像史蒂夫·乔布斯那句我认为苹果每位新员工都听过的话：你意识到周围的一切都是由并不比你更聪明或更有能力的人建造的。我在 Quora 的同事都很出色，但想到这个我青少年时期经常使用的网站是由大约一百人的团队建造的，这很疯狂。这是一次非常赋能的经历。然后我去了麻省理工学院，开始训练自己的神经网络，剩下的就是历史了。

I grew up in Los Alamos, New Mexico. Parents were physicists at the lab. A lot of physicists at that lab. Then I went to work at Quora for about a year. That was my foray and taste of what technology was like. I worked there when I was 17. It was pretty eye-opening in the sense that you really get the Steve Jobs quote that I think every new employee at Apple hears: you realize that everything around you is built by people no smarter or more capable than yourself. My colleagues at Quora were brilliant, but it was crazy to think that this was a site I used to spend a lot of time on as a teenager, and it was built by a team of about a hundred people. It was a very empowering experience. Then I went to MIT, started training neural networks of my own, and the rest is history.

那么你去了麻省理工学院，有点厌倦了学术的学习方面，想成为该领域的实践者。这样说对吗？

And so you went to MIT and you sort of got bored with the learning aspect of academia and wanted to go be a practitioner in the field. Is that fair?

是的，我认为有一件事一直伴随着我：在 2016 年我创办 Scale 时，这一点已经显现出来：要完全实现 AI 或随着时间的推移看到 AI 的成熟，所需的资源将远远超过学术界所能提供的。显然，现在这已经达到了几乎荒谬的程度，数亿、数十亿美元用于训练模型，但这可能是关键驱动力。

Yeah, I think one thing that stuck with me is that it was already playing out at that point in 2016 when I started Scale: it was pretty clear that the amount of resources you would need to fully accomplish AI or to see AI through the fullness of time were going to vastly exceed what was available in academia. Obviously that's true at an almost ridiculous degree now, with hundreds of millions, billions of dollars being used to train the models, but that was probably the key driver.

你从亚马逊将运营和技术结合以实现规模化的做法中获得了哪些启发？

What inspiration have you taken from Amazon with operations and technology being combined for scale?

非常多。我认为亚马逊在很多方面是世界上最反文化的科技公司之一。他们的一个关键洞察是，卓越运营实际上是技术盈余和技术价值的巨大驱动力。杰夫·威尔基（Jeff Wilkey）在那里负责消费者或运营多年，曾任全球消费者业务 CEO（即 AWS 之外的所有业务），他是我的非常亲密的导师。你很快就能学到一种在其他任何科技公司都看不到的思维方式：深度拥抱运营复杂性，将运营视为一门学科；深度拥抱技术与运营的结合，以产生独特强大和有能力的东西；以及极其务实的商业决策方法。这些结合在一起，创造了我们这个时代最伟大的经济引擎之一。所以我学到了很多，Scale 所做的很多事情都是采用同样的方法、剧本和理念：如何将运营复杂性与基础技术突破结合起来，推动整个行业向前发展？

Huge amount. I think Amazon in many ways is one of the most countercultural tech companies in the world. One of their key insights is that operational excellence is actually a huge driver of tech surplus and tech value. Jeff Wilkey, who ran consumer or operations there for many years and was CEO of Worldwide Consumer, so everything outside of AWS, is a very close mentor of mine. You learn pretty quickly that there is a way of thinking there that you just don't see at any other tech company: a deep embrace of operational complexity and operations as a discipline, a deep embrace of the marriage of technology and operations to produce things that are uniquely powerful and capable, and an extremely pragmatic approach to business decision-making. Those in combination created one of the greater economic engines of our time. So I learned a lot, and a lot of what we do at Scale is taking that same approach, playbook, and philosophy: how do you marry operational complexity with fundamental technology breakthroughs to drive an entire industry forward?

亚马逊在并行执行方面也堪称典范。你们在跨产品套件执行时也会考虑这一点吗？

And Amazon's also been kind of canonical in parallel execution as well. Is that something that you guys think about when executing across the suite of different products you offer?

是的。这个洞察的关键美妙之处在于，你弄清楚如何构建问题，使得依赖尽可能少。这样你就可以同时并行下注很多事情。这是投资者非常理解的。如果你有足够多的独立赌注，那么你就可以在那些成功的赌注上加倍下注，最终效果很好。

Yeah. The key beauty of that insight is that you figure out how to architect problems such that you have as few dependencies as possible. So you have as many things you can bet on in parallel at once. That's something investors understand quite well. If you have enough independent bets, then you can double down on the ones that work out, and it ends up working quite well.

你能谈谈吗？我听说你提到过一种二分法：企业因可预测性而获得回报，但实际上却从随机发现中受益。也许以亚马逊为例。

Can you talk a little bit about I've heard you reference the dichotomy of how businesses are rewarded for predictability but actually benefit from elements of random discovery. Maybe using Amazon as an example.

所以，如果你想想亚马逊这家公司，它最初是一家在线书店，然后成了线上万能商店。他们推出了 Prime 会员计划，后来成了全球最大的数据中心提供商。最后这一点听起来很不连贯，就像蹩脚作者写进书里的情节：你有了万能商店，他们规模巨大，然后他们运营了全球所有的计算机。这听起来难以置信。现在，如果你看亚马逊的市值，大多数分析师将公司价值的绝大部分归因于 AWS。所以这个完全不可预测的事件——亚马逊发明 AWS 并建立这项业务——实际上是今天其市值和价值的核心驱动力。这是一个相当疯狂的想法，因为大多数成长型投资者试图直接理解未来几年的收入情况。增长的可预测性如何？但对收益影响最大的却是 AWS 被发明这个完全不可预测的事件。所以公司有一个令人困惑的特性：一方面，投资者认为他们在押注公司未来几年的执行，但对于最好的公司，他们实际上押注的是持续的重塑。我认为英伟达是最好的现代例子。英伟达是一家销售游戏和图形芯片的 GPU 公司，持续了几十年。大约 15 年前，他们注意到人们开始使用英伟达 GPU 训练 AI 算法，因为并行计算能力。他们开始投入大量时间、精力和研发来支持这个用例，这需要早期对 AI 有很强的信念，并在财务上产生重大影响之前持续投入。但今天，英伟达是一家万亿美元公司，几乎纯粹因为 AI。所以如果你 10 年前投资英伟达股票，你评估的是他们在图形和游戏芯片上的执行能力，但真正决定你是否能赚大钱的是他们是否将自己重塑为一家 AI 公司。我认为这是市场和公司的核心，很多人不理解：押注几乎总是押在重塑能力上。你如何在 Scale 内部文化中体现这一点？你们曾经是 Scale API，后来 Scale AI 主要专注于自动驾驶。现在范围更广了，做 RLHF 相关的事情。听起来这是你研究和思考的东西。你如何确保这在公司文化中存在？

So, if you think about Amazon as a company, it was an online bookstore, then the everything store online. They created Prime, a membership program, and then it became the largest data center provider in the world. That last piece sounds so non sequitur, like a bad author would write into a book: you had the everything store, they were so big, and then they ran all the computers globally. It sounds unbelievable. Now, if you look at Amazon's market cap, most analysts attribute the vast majority of the company's value to AWS. So this very unpredictable event—Amazon inventing AWS and building that business—is actually the core driver of its market cap and value today. It's a pretty crazy thought because most growth investors try to directly understand what will happen to revenue over the next few years. How predictable is growth? But the thing that affected earnings the most was this totally unpredictable event of AWS being invented. So there's this confusing property of companies: on one hand, investors think they're betting on the next few years of execution, but for the best companies, they're really betting on continuous reinvention. I think Nvidia is the best modern example. Nvidia was a GPU company selling gaming and graphics chips for decades. About 15 years ago, they noticed people using Nvidia GPUs to train AI algorithms because of parallel computing capability. They started investing huge amounts of time, effort, and R&D into supporting that use case, requiring a lot of conviction in AI to start that investment early and keep leaning into it long before it moved the needle financially. But today, Nvidia is a trillion-dollar company almost purely because of AI. So if you were an investor in Nvidia stock 10 years ago, you were evaluating their ability to execute on graphics and gaming chips, but the thing that actually matters for making a ton of money is whether they reinvent themselves as an AI company. I think this is the core of markets and companies that many don't understand: the bet is almost always on the capability to reinvent. How do you manifest that culturally within Scale? You guys were Scale API once upon a time, then Scale AI focused mostly on autonomous vehicles. Now it's much broader, doing stuff around RLHF. It sounds like this is something you study and think about. How do you make sure that exists culturally within the business?

好问题。这就是为什么我花很多时间思考这个问题。我们做了几件事。我认为任何时候都有更多可以做的。一是我们尽可能创造一种纯粹的精英管理文化，这种文化大力倾向于公司里通常更年轻的员工，他们有好的想法，几乎被赋予责任去推行这些想法并把它们变成大事。这种文化——如果你有一个好主意，首先，任何人都可以有好主意，如果你有一个好主意，你几乎要全权负责实现它并让它发生。这种文化不是大多数公司的运作方式。大多数公司，每个人都可以有好主意，然后某个总监或副总裁偷了你的主意，把它变成自己的职业跳板。这种文化相当独特，我们大力推行，明确表明你在 Scale 的影响力和未来的真正限制是无限的，取决于你投入多少、你的想法有多好、有多创新等等。这是其一。另一个是，我们总是努力让自己专注于大问题，如果合理的话。很多时候，亚马逊的版本是专注于客户，但如果你在系统中有一个正确的固定点——对亚马逊来说是客户，对我们来说是思考行业中的大问题——那么你最终总会偶然发现越来越大的机会。我的意思是，我们很长时间专注于自动驾驶，这是一个巨大、复杂、有趣的问题。在某个时刻，我们之前讨论的地缘政治以及 AI 对未来国家间力量平衡的重要性变得非常清晰。我们对此有很强的信念。我们大力投入与美国政府和国防部的合作，我们在服务自动驾驶行业时积累的很多技术也相当适用。但我们随后接手了一个更大的问题：如何确保美国的领导地位，如何确保美国保持领先。这个问题如此之大，以至于在服务这个问题的过程中，我们偶然发现了比自动驾驶最初机会大得多的机会。现在同样如此，大问题是帮助确保美国在 AI 行业取得最大进展。我们如何确保这些模型成为最具影响力的版本，推动 AI 行业取得最大进步？这是我们这个时代最大的问题。

Great question. It's why I spend a lot of time thinking about it. There are a few things we do. I think there's certainly a lot more you can do at all times. One is that we create a culture of pure meritocracy as much as possible, one that leans heavily into people who are usually more junior at the company who have good ideas being almost thrown into the responsibility of having to run with those ideas and turn them into something big. This culture—if you have a great idea, first of all, anyone can have a great idea, and if you have a great idea, you have almost full accountability to realize it and make it happen. This kind of culture is not how most companies operate. Most companies, everyone can have a good idea, and then some director or VP steals your idea and makes it their career move. This culture is pretty unique, and we really lean hard into it to make it very clear that the true limit to your impact and future at Scale is limitless, depending on how much you apply yourself, how good your ideas are, how innovative they are, etc. That's one. Another is that we try to always put ourselves to focus on big problems if it makes sense. A lot of times, Amazon's version of this is focusing on the customer, but if you have the right fixed point in the system—for Amazon it's the customer, for us it's thinking about the big problems in the industry—then you'll always end up stumbling upon opportunities that continue to get bigger and bigger. By that I mean, we were focused on autonomous vehicles for a very long time, which is a huge, complicated, interesting problem. At a certain point, it became pretty clear that a lot of what we talked about with geopolitics and the importance of AI to the future balance of power between countries was going to be the case. We had high conviction in that. We leaned very hard into working with the US government and the US DoD, and a lot of the technology we built up in servicing the autonomous vehicle industry was pretty applicable. But we then took on this much larger problem of how to ensure American leadership and how to ensure the US stays ahead. That's such a big problem that in the course of serving that problem, we stumbled upon much larger opportunities than the original ones in autonomous vehicles. The same has been true now with the big problem of helping to ensure maximal progress for the US in the AI industry. How do we ensure these models are the most impactful versions of themselves, pushing for the maximum amount of progress in the AI industry? That's the biggest problem of our time.

我想聊聊面试。你说你最喜欢的面试问题是‘你曾经在什么事情上最努力过？’你为什么喜欢这个问题？

I want to ask about interviewing. So, you said your favorite interview question is, 'What's the hardest you've ever worked on something?' Why do you like that question?

是的。我通常认为世界上有两种人。有一个心理学概念：内控型和外控型。如果你有内控型，意味着你相信生活中发生的事情更多是你自己行为和行动的结果。所以你相信你掌握着自己人生的缰绳。如果是外控型，则相反。你相信发生在你身上的事情大多是外部因素的结果，世界很宿命，你就像弹球机里的一颗弹球。我很喜欢这一点，如果你知道如何观察，这是人们看待自己生活的一个非常清晰的分野。我发现我只想和那些有内控型的人共事。一个观察方法是看人们在对自己重要的事情上有多努力。每个人都有对自己重要的事。但如果他们是内控型，他们会拼命工作确保那些事情以最好的方式发生。如果是外控型，事情对他们也重要，但他们基本上会摊手放弃，让世界来掌控。所以通过看人们在最重要的事情上有多努力，量化他们有多执着、多在乎、多关注细节，你就能清楚了解他们对自己人生结果有多少掌控感。

Yeah. So, I generally think there are really two kinds of people in the world. There's a psychological term: having an internal versus an external locus of control. If you have an internal locus of control, it means you believe the things that happen in your life are more a product of what you do and the actions you take. So you believe you're holding the reins on your own life. If you have an external locus of control, it's the opposite. You believe the things that happen to you are mostly outcomes of things outside your control, like the world is very deterministic and you're a pinball in a big pinball machine. I really like that if you know how to look for it, this is a very clear dichotomy between how people think about their lives. I find that I only want to work with people who have an internal locus of control. One way to look at that is seeing how hard people work at things that matter to them. There are things that matter to everybody. But if they have an internal locus of control, they'll work their ass off to make sure those things happen the best possible way. If they have an external locus of control, things matter to them, but they sort of throw their hands up and let the world take the wheel. So by seeing how hard people work on the things that matter most to them, quantifying how obsessive they were, how much they really care, how small of details they sweat, you get a clear indication of how much control they believe they have over their life outcomes.

你在招聘中最看重的单一特质是什么？是那种控制点，还是其他什么更突出的？

What's the single trait characteristic that you're most looking for in hiring? Is it that locus of control or is there something else that stands out?

是的，我认为有几个。我们早期有一份文档，写的是我们在招聘中看重什么，大概有四个特质。第一是内控型。第二是问题解决者。从根本上说，就是非常擅长创造性解决问题的人。你给他们一个问题，他们能想办法绕过障碍。这是非常重要的特质。第三，我们寻找令人印象深刻的人。当你和他们交谈、共事时，你会由衷地对他们印象深刻。这代表那些不断提升组织标准的人。如果你对某人印象深刻，你会很有动力每天来上班，和他们一起工作并向他们学习。我们在这方面标准很高。最后一个是协作精神。你可能遇到内控型强、善于解决问题、令人印象深刻，但就是很难共事的人。所以这些就是组织的北极星，它们让我们走得很远。

Yeah, I think there are a few. We had an early document we wrote up around what we look for in people we hire, and there were sort of four traits. One is internal locus of control. Two is problem solvers. Fundamentally, people who are very good at creative problem solving. You give them a problem and they figure out a way around the roadblock. That's a really important trait. Third, we looked for people who are impressive. When you talked with them and worked with them, you were genuinely impressed by them. It's a shorthand for people who are constantly upping the bar of the organization. If you're impressed by somebody, you're very motivated to come to work every day, work with them, and learn from them. We held a pretty high bar there. The last one was people were collaborative. You can have people who are high locus of control, good problem solvers, very impressive, but just suck to work with. So those were the north stars for the organization and carried us pretty far.

你谈到过大品牌和科技公司的声望实际上扭曲了硅谷的招聘视角。你怎么看？比如人们长期待在这些大品牌公司。为什么这反而是一个反向信号？

You've spoken about how the prestige around big brands and tech actually perverts and distorts the perspective around hiring in Silicon Valley. How do you think that's the case? Like these big brand names people stay at for a long time. Why is that kind of a contra signal?

关于这一点，我最喜欢的一句话是：如果你的招聘组织看起来像大学招生办公室，那你应该很害怕。我认为这是真的。现实是，在大科技公司很难有真正的影响力。我不是在指责大科技公司，但他们招了太多人，真正重要的问题范围有限。所以很多被招进去的人最终只做问题的一小部分中的一小部分中的一小部分。如果你考虑选择偏差，那些被选入这些大品牌科技公司的人，是那些相对于影响力过度优化品牌和地位的人。相比之下，小型初创公司则完全相反。你加入一家小型初创公司，是因为你看到五个人在做这件事，你知道自己可以进来产生很大影响。但这不会是一件很酷的事；你不能告诉朋友你在这家初创公司工作，让他们觉得这很了不起。所以很多招聘是基于技能的，但也有很多是文化上的测试。你真正想要的是那些不在乎地位、非常在乎影响力的人。我认为大科技公司在这方面是负向选择的。

One of my favorite lines around this is: if your recruiting organization looks like a college admissions office, then you should be pretty scared. I think it's true. The reality is it's very hard for somebody at a big tech company to have any sort of real impact. This is not too much of an indictment of big tech companies, but they just hire so many people, they have a limited scope of problems that really matter. So a lot of the people they hire end up working on a teeny piece of a teeny piece of a teeny piece of a problem. If you think about the selection bias, the people who get selected into these very large brand name tech companies are those who are overoptimizing for brand and status relative to impact. By contrast, small startups are the exact opposite. You join a small startup because you see the five people working on this thing and you know you can come in and have a big impact. But it's not going to be a cool thing; you won't be able to tell your friends about working at this startup and have them think it's awesome. So a lot of hiring is skills-based, but a lot of it is also culturally testing people. You really want people who don't care about status and care a lot about impact. I think big tech companies negatively select for that.

我们之前聊过《从 0 到 1》，以及它如何在创业文化中变得常态化。曾经它非常革命性，但现在书里的很多内容已经成为常态。你最近有没有读到或内化一些关于创业的非共识观点，你认为在未来几年内会变得重要，关于如何运营或与公司合作？

We were talking about zero to one before we got going and how it's kind of been normalized in startup culture. Once upon a time it was very revolutionary, but now a lot of the things in that book have become status quo. Is there something you've read or internalized recently about startups that's non-consensus that you think will be at some point in the next couple years around how to operate or work with companies?

我认为有一件事在生态系统中肯定是非共识的，但却是我的经验：那些非常努力但未必在你公司有丰富经验的人的价值。这相当令人惊讶。

I think one thing that is certainly non-consensus in the context of the ecosystem but has been my experience is the value of very hardworking people who are not necessarily super experienced in your company. It's pretty surprising.

有些公司是由一小群非常有经验的人打造出不可思议的东西，这确实存在。但大多数情况下，我认为大多数初创公司就像一群混乱忙碌的人，他们不一定非常有经验，但非常勤奋、能力很强、很有才干。他们几乎像梯度下降一样构建出这些令人惊叹的东西。我认为这一点并没有被整个科技生态系统很好地理解或采纳。很多科技生态系统非常注重招聘那些有经验、经历过的人。

There are some companies where a small group of very experienced people build something incredible. That certainly exists. But for the most part, I think most startups are a chaotic buzz or hive of people who are not necessarily super experienced, but very hardworking, very high aptitude, and very capable. They sort of almost like gradient descent to building these incredible things. I think that's not super well understood or adopted by the entire tech ecosystem. A lot of the tech ecosystem is really focused on hiring experienced people who have been there and done that.

另一件事，我们稍微讨论过，就是拥有强烈观点的重要性。很有意思：上一代科技巨头——想想谷歌、Meta，甚至苹果——创业建议或经典商业建议是尽可能保持观点中立和品牌中立，这样你可以广泛分发产品，不得罪任何人，并产生尽可能广泛的影响和吸引力。我认为我们正在迅速进入一个非常不同的时代，正确的做法是拥有相当强烈的观点，并大声宣扬这个观点，因为这能吸引认同你的人才。这对于建立积极的文化和非常高素质的团队非常有益。这对你的客户也非常重要，因为越来越多的客户——无论是企业还是消费者——非常看重与在理念上认同他们、分享他们观点的人合作。这迫使你保持公司的真实性。这是一个微妙的事情，但我看到很多企业软件领域的同行，这些公司很快就变得毫无立场。早期，每家公司都是创始人的产物，他们非常在意，对每个细节都倾注心血。然后无一例外，每家企业软件公司都变成了工具包里的另一个小部件。我认为公司保持身份认同和真实至关重要，这样才能有机会实现我之前提到的重塑。

The other thing, and we're talking a little bit about this, is the importance of having a strong point of view. It's quite interesting: the last generation of tech giants—you think about the Googles, the Metas, and even the Apples of the world—the startup advice or classic business advice is to have as neutral a point of view and as neutral a brand as possible, so that you can distribute a product broadly, not offend anyone, and have as wide-scale impact and broad-based appeal as possible. I think we're very quickly entering a very different era, where the right thing to do is to have a pretty strong point of view and to be very loud about that point of view, because that allows you to attract the talent of people who agree with you. So it's incredible for building a positive culture and a very high-talent group. It's also very important for your customers, because more and more customers—whether enterprise or consumer—care a lot about working with people who philosophically agree with them and share their points of view. And it forces you to keep your company authentic. It's a subtle thing, but I look at a lot of peers in enterprise software, and these companies very quickly stand for nothing. Early on, every company was the product of founders who cared a lot and sweated every detail. Then invariably, every enterprise company becomes another widget in the bag of tools. I think it's important for companies to maintain a sense of identity and stay authentic to have any chance at the reinvention component I talked about before.

Scale 从来都不是一个特别酷的业务。你能详细说明一下吗？这对公司来说是净负面还是净正面？

Scale has never been a particularly cool business. Can you elaborate on that and if that's been a net negative or a net positive for the company over the years?

完全同意。我觉得这很有趣。我们一直在非常酷的领域运营——自动驾驶汽车、当前的人工智能革命——但我们从来不是这些领域里酷的人，因为从根本上说，我们是基础设施提供商。基础设施并不那么性感。实际上，对于我们的公司，我们不想要那些只想耍酷、花哨、从事令人兴奋的新技术的人。我们真正想要的是那些愿意卷起袖子、弄脏双手、解决人工智能中那些不性感但非常重要的问题的人。所以我认为这对于以符合我们需要完成的工作的方式建立公司非常重要。其影响是，加入 Scale 的人知道他们将要面对什么。他们知道我们在生态系统中扮演的角色，并且他们非常在意这一点。

Totally. I think it's funny. We've always operated in very cool spaces—self-driving cars, the current AI revolution—but we've never been the cool people in those spaces, because fundamentally we're an infrastructure provider. Infrastructure is not that sexy. Actually, for our company, we don't want the people who just want to be cool and flashy and work on exciting new technologies. We really want the people who are willing to roll up their sleeves, get their hands dirty, and work on the unsexy problems in AI that are really damn important. So I think it's been very important for building the company in a way that is true to the work we need to do. The impact has been that the people who join Scale know what they're getting into. They know what role in the ecosystem we play, and they care a lot about that.

你现在坐的办公室很棒。我听说你认为应该花钱在好的办公空间上，这对员工很重要。你能谈谈为什么吗？

You have a wonderful office that we're sitting in right now. I've heard you say that you believe you actually should spend money on a nice office space and that it's an important thing to do for your employees. Can you talk to me a little about why that's the case?

是的，冒着听起来有点玄乎的风险，我确实认为你所在的空间对你的思维有很大影响。就我个人而言，待在自然光充足的空间里是我能为思维质量做的最好的事情之一。我认为有很多分形效应，空间质量的细微差别、自然光的多少，或者与同事的配置方式，都会对最终结果和思维质量产生相当大的影响。所以这是那种影响程度几乎难以察觉、而且有点反直觉的事情。

Yeah, at the risk of sounding somewhat woo-woo, I do think that the spaces you're in impact a lot about your thinking. Personally, being in spaces with a lot of natural light is one of the best things I can do for the quality of my thinking. I think there are a lot of fractal effects where pretty subtle differences in the quality of your space, the amount of natural light, or the configuration with your co-workers can have pretty big impacts on the ultimate end outcome and the quality of thought. So it's one of these things that is almost insidious in how much it matters, and sort of unintuitive.

那如何安排你的一天呢？你如何安排一天以实现最大生产力？

What about structuring your day? How do you structure your day for maximum productivity?

我经常发现最好的做法是在一天开始时设定一些非常明确的目标，问自己：‘今天对我来说最重要的事情是什么？’它们可以从很小的事情开始，然后随着时间的推移，你会找到自己的极限并扩大目标。然后我每天有很多会议，这是工作的一部分，但我会不断检查自己在实现设定目标方面的进展。我认为这可能是最好的做法。

What I often find is the best thing to do is to set some pretty clear goals at the very start of the day, to say, 'What are the most important things for me to get done today?' They can start out pretty small, and then over time you'll find what your limits are and upsize them. Then I'm in a ton of meetings every day, which is part of the job, but I'll continually check in on how I'm progressing against the clear goals I set. I think that's probably the best thing I would do.

我听说你小时候数学和物理都有明确的正确答案，但小提琴对你影响很大，因为它不仅仅是把音符拉对。你能详细说明一下，或者谈谈小提琴为什么以及如何影响了你吗？

I've heard you say that both math and physics growing up had clear right answers, but it was violin that was super influential to you because it wasn't just about getting the notes right. Can you elaborate on that point or expand on why and how the violin influenced you?

我认为对于在商业中非常量化的人来说，令人有些抓狂的是，你总是在一个灰色地带运作，因为你永远无法真正知道你的决定是完全正确还是错误。大多数重要的事情都很难衡量，你只能凭直觉行事。所以这种模糊思维或直觉驱动的思维在数学和科学中并没有得到很好的训练，但在美国的艺术中却得到了更好的训练。所以这是它对我形成影响的主要方式。另一件非常重要或有价值的事情是培养品味。我认为公司的产品很大程度上是品味的产物，以及你对这种品味重视的程度。

I think one of the things that is somewhat maddening for people who are very quantitative in business is that you're constantly operating in a bit of a gray area, in the sense that you'll never really know if your decisions were fully correct or incorrect. Most things that matter are quite hard to measure, and you just have to operate via instinct. So this sort of fuzzy thinking or intuition-driven thinking is not super well trained in math and science, but much more well trained in the arts in America. So that's the primary way it's been formative. Another thing that's been quite important or valuable as part of that is also developing a sense of taste. I think so much of the product of a company is an outcome of taste and the degree to which you take that taste seriously.

对人的品味、对美学的品味、对产品的品味、对组织方式的品味。苹果可能是最好的例子，是世界上最有品味的公司之一。对我来说，身处一个需要培养品味才能有效并将其应用到公司中的领域一直很重要。

Taste in people, taste in aesthetics, taste in product, taste in how to organize. Apple is probably the best example of this, one of the most tasteful companies in the world. It's been important to me to have been in a field where you have to develop taste to be effective and apply that to the company.

你父亲是物理学家，母亲是天体物理学家，对吗？你的童年对今天的 CEO 兼创始人 Alexandr Wang 影响最大的是什么？

Your dad's a physicist and mom's an astrophysicist, right? How did your childhood most influence the CEO and founder that Alexandr Wang is today?

这方面有个很好的例子。我刚刚和父母一起过了周末。对他们来说，重要的是他们共事的人和领导对工作场所、工作本身以及领域历史有着非常深刻、几乎无法解释的热情。同样，我父母都看了很多遍《奥本海默》。他们告诉我，我们必须反复看，因为要弄清楚所有物理学家是谁——电影里那些只有一句台词或没有台词的物理学家。我们必须搞清楚每个物理学家是谁演的。所以我父母对物理学有着这种无法解释的热情，这种对领域根本性的关心和热爱深深影响了我。我妈妈从我出生起就教我物理学。那种深层的热情非常有效。

We have a great example for this. I just spent the weekend with my parents. To them, it was really important that the people they worked with and their leaders had this very deep, almost inexplicable passion for the place, the work, the history of the field. Similarly, my parents both watched Oppenheimer many times. They told me we had to keep rewatching it because we had to figure out who all the physicists were—physicists in the movie that had a single line or didn't have any lines. We had to really figure out who played each physicist. So there's this level of inexplicable passion for physics that both my parents have, this fundamental care and love of the field that really rubbed off on me. My mom taught me about physics ever since I was born. That level of deep enthusiasm has been quite effective.

你写过一篇博客文章《雇佣那些在乎的人……》，我觉得这和我们之前谈到的招聘问题有关。在招聘过程中，你还有什么其他方法来判断某人是否对你的公司有独特的热情，而不是对其他任何企业？

You wrote a blog post 'Hire people that give a...' I think that ties into some of the hiring things we were speaking about earlier. Is there anything else you would say about how you try to assess if someone uniquely has the passion for your company versus any other business when recruiting?

我们常做的一件事是问人们为什么来 Scale 面试。你可以从答案的独特程度判断好坏。如果人们只是说‘哦，AI 是下一件大事，我想在 AI 公司工作’，那还行。但如果他们说‘我和一个朋友一起训练模型，我们花了五个小时看数据。数据里有一个小错误导致整个模型无法工作，我意识到这个问题非常有趣，然后我申请了 Scale，因为这才是正确的答案。’所以你要寻找的——Paul Graham 对此有很精辟的论述——是人们关心某事的理由，当这种关心是非理性的时候，无论是出于好奇心还是某种怪癖。某种根本非理性的原因，让他们关心我们所做的事。这可能是我们最看重的：一些根本非理性、难以解释的热情。

One thing we do is often ask people why they're interviewing at Scale. You can tell a good answer by how obscure it is. If people just say, 'Oh, AI is the next big thing and I want to work in an AI company,' that's okay. But if they say, 'I was working with one of my friends to train a model, and we spent five hours just looking at the data. There was one little bug in the data that caused the whole model to not work, and I realized that problem was really deeply interesting, and then I applied to Scale because that's the right kind of answer.' So one of the things you look for—and Paul Graham has written very elegantly about this—is a reason for people to care about things when there's an irrational reason to care, whether because of some curiosity or quirk. Something fundamentally irrational, some reason they care about what we do. That's probably the thing we look for most: something that is fundamentally irrational and hard to explain about their passions.

就像音乐，对吧？练习音乐，也许别人永远不会知道你省掉了最后一个细节。但如果你知道，如果你真的练习了，那它就变成了你内在的东西。

Similar to music, right? Practicing music, maybe people will never know if you cut that last corner. But if you know it, if you really practice it, then it's something that's innate to you.

完全同意。

Totally.

我听过你说，也许你在推特上说过，你一生都很古怪，你尊重过的每个人也都很古怪。为什么你认为古怪是成为一个有趣的人以及与你共鸣的人的重要特质？

I've heard you say, maybe you tweeted or something, you've been weird your whole life and that everybody you've ever respected has also been weird. Why do you think being weird is an important trait to being an interesting person and the types of people you resonate with?

纯粹从统计上看，如果你正常，意味着你在钟形曲线内，而在钟形曲线内很难取得伟大成就或对世界产生巨大的差异化影响。所以这是一个纯粹的统计论点。但我认为最有趣的是，正常大致意味着拥有主流信念。这没什么错，但这意味着如果你正常，模拟与你的对话相当容易。在某种程度上，这样的对话信息量很低。而如果你古怪，说很多非常出乎意料的话，有很多意想不到的想法，那会是一种非常有生成性的体验。所以与古怪的人互动、让自己被古怪的人包围最终非常有价值，因为你可以在一个更熵增、更根本上有趣和多样化的思想和观念池中沐浴。这是你能拥有的最大礼物。

Purely statistically, if you're normal, that means you're in the bell curve, and it's hard to be in the bell curve and accomplish great things or have a huge amount of differentiated impact on the world. So it's a pure statistical argument. But I think the thing I find most interesting is that being normal is kind of an approximation for having generally mainstream beliefs. There's nothing wrong with that, but it means that if you're normal, it's pretty easy to simulate a conversation with you. In some ways, there's low information content from having that conversation. Whereas if you're weird and you say a lot of very unexpected things and have a lot of unexpected thoughts, that's a very generative experience. So interacting with people, surrounding yourself with weird people ends up being quite valuable because you get to bathe in a more entropic and more fundamentally interesting and diverse pool of ideas and thoughts. That's the greatest gift you could have.

展望未来五到十年，AI 的什么最让你兴奋？

What has you most excited about the future of AI as we look out five to ten years from now?

很难不为我说过的感到兴奋：这可能是人类有史以来最伟大的经济发明和最强大的经济引擎。这从根本上令人难以置信地兴奋。就像发明了蒸汽机的一百万倍。这个东西将产生巨大的经济盈余，让许多人过上更好的生活，将人类提升到如此疯狂的程度。深入来看，最令人兴奋的部分是提升人类状况。以医疗保健为例。全球范围内，医生短缺大约十倍。因为培训需要大量时间和资源，医生太少了。即使有这些医生，医疗保健大多是被动反应的。

It's hard not to be excited about what I talked about: potentially the greatest economic invention and the greatest economic engine that humanity will have ever invented. It's fundamentally incredibly exciting. It's like inventing the steam engine times a million. So this thing will generate so much economic surplus that lifts so many people into better living conditions, elevating humanity to such an insane degree. Double-clicking on that, the deeply exciting components are the elevation of the human condition. Take healthcare, for example. Globally, there's roughly a 10x shortage of doctors. Because it takes so much training and is so expensive to train people, there are way too few doctors. And even with those doctors, healthcare mostly works reactively.

就像你去看医生，你有个问题，有时他们能轻松解决，有时治疗极其昂贵，大多数时候解决起来非常贵，而且有时还治不好。你知道，我们从根本上需要一个更主动的医疗系统，通过持续测量很多东西，你可以在非常早期就处理这些问题。医疗是一个完整的领域，没有技术突破，我们作为物种就卡住了。人类在如何在没有真正根本性技术进步的情况下改善医疗方面，有点停滞不前。所以如果 AI 突然能给每个人一个口袋里的医生，让他们一感觉到奇怪或觉得不对劲，或者有个奇怪的肿块什么的，就能主动应对，那真是太棒了。这只是其中一种方式，它可能对我们所做的任何事情中，对寿命、全球寿命产生最大的影响之一。所以这些是让我非常兴奋的事情，全面的连锁反应将会非常棒。

Like you go to the doctor, you have a problem, and sometimes they can fix it easily, sometimes it's extremely expensive to fix, most of the time it's very expensive to resolve, and then sometimes it doesn't work out. You know, fundamentally we need a more proactive healthcare system by which you're constantly measuring a lot of things and you can deal with these problems very early. And healthcare is an entire field that, without technology breakthroughs, we're kind of stuck as a species. Humanity is a little bit stuck in how good you can make healthcare without real fundamental technological advances. So if AI all of a sudden can give everybody a doctor in their pocket that enables them, as soon as they feel something weird or they think something weird's going on, or there's a weird bump or whatever, they can be proactive about that. It's pretty incredible. That's just one way in which that could have one of the greatest effects on longevity of anything that we do, global lifespan. So those are the things that get me really excited, the full knock-on impacts are going to be pretty great.

Alex，感谢你接受采访。

Alex, thanks for doing this.

嗯，非常感谢你邀请我。

Yeah, thanks so much for having me.