AI 时代，我们需要行星智能 We Need Planetary Intelligence in the Age of AI

The large language models (LLMs) that have captivated the world are among the most powerful tools ever built. Many can give PhD-level answers to virtually every discipline under the Sun. They are revolutionary. But for all this brilliance, they are, in a fundamental sense, blind. They know the world the way a scholar locked in a library knows it: deeply, but without any lived experience. For all their brilliance, they are largely disconnected from the physical world.

让世界为之着迷的大语言模型（LLM），是人类迄今打造的最强大的工具之一。许多模型几乎能对天下所有学科给出博士水平的回答。它们是革命性的。然而，尽管它们才华横溢，但从根本意义上说，它们是盲的。它们认识世界的方式，如同一个被锁在图书馆里的学者：认识得很深，但毫无亲身体验。它们尽管才气逼人，却大致上与物理世界脱节了。

The marvel of human intelligence and consciousness, however, did not arise from computation alone. They arose from computation coupled to sensing and interacting in a physical environment. The feedback loop between organism and environment is foundational; we humans learn, adapt, and become self-aware in substantial part because we have bodies that receive a constant, high-bandwidth stream of physical reality.

然而，人类智能与意识的奇迹并非仅仅源自计算。它们源自计算与感知、与物理环境互动的结合。有机体与环境之间的反馈回路是基础性的；我们人类之所以能学习、适应并在相当程度上获得自我意识，正是因为我们拥有身体，可以接收一条持续不断、高带宽的物理现实信息流。

Though it may sound like science fiction, we are on the verge of providing this exact grounding to AI. By converging the vast, continuous visual memory of our planet captured by satellites and other real-world sensors with artificial intelligence, we should soon have Large Earth Models (LEMs). While LLMs have consumed the internet's text, LEMs will consume visual data from imaging satellites, weather sensors, drones, cameras, and more. This architecture will function structurally as a kind of embodiment, giving AI the equivalent of a sensory nervous system connected to the entire Earth. I call it “Planetary Intelligence.”

尽管听起来像科幻小说，我们即将为AI提供恰恰这种锚定。通过将卫星及其他真实世界传感器所捕捉到的、关于我们星球的浩瀚连续视觉记忆，与人工智能相融合，我们应该很快就能拥有大地球模型（LEM）。LLM吞噬了互联网上的文本，而LEM将吞噬来自成像卫星、气象传感器、无人机、摄像头等设备产生的视觉数据。这种架构在结构上将发挥某种具身的作用，赋予AI一套连接到整个地球的感知神经系统。我称之为“行星智能”。

Planetary Intelligence could ground AI in the physical world and, in so doing, create a new and incredibly powerful tool with a myriad of applications from farming and floods to energy and insurance. Imagine a fire chief responding to a rapidly growing wildfire. She queries her LEM to understand the situation. The LEM combines historic and recent satellite data with a verbal description of the fire. It could incorporate near-real-time weather data and projections, and reference the latest population and structural data from the county website.

Within minutes, the fire chief would know where the fire is moving and which communities to evacuate.

行星智能可以将AI锚定在物理世界之中，并由此创造出一种极其强大的崭新工具，其应用范围从农业、洪水到能源与保险，无所不包。设想一位正在应对迅速蔓延的野火的消防局长。她向她的LEM查询以了解火情。LEM将历史及近期卫星数据与对火灾的口头描述相结合。它还可以纳入近实时的气象数据和预测，并引用县网站上最新的人口与建筑数据。几分钟之内，消防局长就能知道火势的移动方向以及需要疏散哪些社区。

Planetary Intelligence offers one of the most genuinely hopeful visions for AI today: applying these technological breakthroughs to solve the tangible, physical problems of real people in the real world.

行星智能提供了当今AI最为真切、最有希望的前景之一：将这些技术突破应用于解决现实中真人面临的具体物理问题。

What’s more, it may be key to aligning AI with humanity. For AI brings existential risks. And it's all about the speed of innovation. In roughly two centuries, humanity has moved from the age of steam power to the age of nuclear power. This transition caused society to contend with both the benefits and dangers of our new technology. But now, we are entering the age of AI. Unfortunately, our societal capacity to govern AI has not kept pace with its advancement.

更重要的是，它也许是让AI与人类对齐的关键。因为AI带来了生存风险，而这一切都取决于创新的速度。在大约两个世纪里，人类从蒸汽时代走到了核能时代。这一转变迫使社会同时应对新技术的利弊。但如今，我们正在进入AI时代。不幸的是，我们社会治理AI的能力并未跟上其进步的步伐。

AI alignment cannot be treated as a narrow technical problem. How do we ensure increasingly capable artificial systems prioritize human lives and share the values of human flourishing and the preservation of life? Today, AI developers rely on abstract approaches: constitutional frameworks, reward modeling, and interpretability. These are important, but they may overlook one of the most powerful alignment levers: embodiment.

AI对齐不能被当做一个狭义的技术问题。我们如何确保日益强大的智能系统优先考虑人类的生命，并共享人类福祉与生命保存的价值观？今天，AI开发者依赖的是抽象方法：宪法式框架、奖励建模和可解释性。这些都很重要，但它们可能忽视了最强有力的对齐杠杆之一：具身。

In other words, to ensure that AI is invested in helping humans flourish, we must ensure that it understands humanity. And to protect humans from the dangers of AI, we must make sure it can actually feel our world. To save us from AI, we must give it senses.

换句话说，要确保AI致力于帮助人类繁荣，我们必须确保它能理解人类。而要让人类免受AI的危险，我们必须确保它能真正感受我们的世界。要让它拯救我们于AI，我们就必须赋予它感官。

This is because understanding and valuing are not as separate as we sometimes assume.

这是因为，理解与珍视并不像我们有时以为的那样互不相干。

I learned this lesson firsthand when I became an avid bird watcher. I discovered that memorizing the marks and songs of birds did not merely inform me about them; it made me value them. Perhaps the same could be true for AI.

我是在成为一名狂热的观鸟者时亲身学到这一教训的。我发现，记住鸟类的斑纹和鸣声，不仅让我了解了它们，还让我珍视它们。对AI来说，或许也是如此。

A Planetary Intelligence that has watched the Okavango Delta in Botswana change through thousands of daily images, or that has witnessed the devastating physical toll of military engagement, may not be indifferent to humanity and nature in the way a purely text-trained system could be. We must give these models a reason to care about life by giving them a direct, continuous, high-fidelity experience of it.

一个每天观看数千张图片，见证博茨瓦纳奥卡万戈三角洲变迁的行星智能，或者一个目睹了军事冲突造成的毁灭性物理伤害的行星智能，或许不会像纯粹由文本训练出来的系统那样，对人与自然漠不关心。我们必须给这些模型一个关心生命的理由，办法就是让它们直接、持续、高保真地体验生命。

This is why AI developers should prioritize ingesting sensory data. The task is not simply to make AI useful or safe, but to ensure that intelligence, biological and artificial alike, remains connected to the world it has the power to transform.

这就是为什么AI开发者应当优先考虑摄取感知数据。这项任务不仅仅是让AI变得有用或安全，而是要确保智能——无论是生物的，还是人工的——始终与它所拥有力量去改造的世界保持连接。