Planetary Intelligence - Will Marshall
Will Marshall
SubscribeSign in
Planetary Intelligence
Will Marshall<br>May 27, 2026
Share
Everything, everywhere, all at once.
In this essay, I introduce an idea for a new type of machine intelligence that understands our physical world in real-time – a powerful expansion of AI’s capabilities. I talk about implications for people across economic, security, and sustainability domains. And I speculate about humanity's place in the cosmos, and how giving AI sensors may help humans and machines coexist safely and survive the Great Filter.
Photo by NASA
I. The Models That Are Blind
The AI large language models that have captivated the world have consumed the written record of human civilization: every article, essay, book, and conversation that humanity has committed to the internet. From that diet, they have become extraordinarily fluent, knowledgeable, and capable of reasoning across vast domains. They are the most powerful intellectual tools ever built.<br>And yet they are, in a fundamental sense, blind.<br>They know about the world the way a scholar locked in a library knows about it: deeply, but without lived experience. Ask an LLM what is happening in a farmer’s field in Kansas today, and it will offer everything it has absorbed about agronomy and soil science, historic weather patterns and the history of farming in the Midwest, but it cannot tell you what is actually happening on that farm today, and how that compares to yesterday and last year. For all its brilliance, it is disconnected from the physical world. Because AI is only as good as the data it is trained upon and it only has what’s publicly available on the internet.<br>This is not a minor limitation. The physical world is where daily life occurs: the building, the field, the disasters, the conflicts, the biodiversity and ecological destruction, the migrations, the exploration of the oceans. These are not text events. They are events in matter, in soil, in water, in the movement of people and armies and atmosphere.<br>The physical world is where real life unfolds, and until now, AI has had no way of sensing it.
II. Data: A Second Foundation
If the text of the internet, from Wikipedia to Reddit, were core to the initial development of LLMs, what is the equivalent for real world models? I’m biased, but I’ll argue it’s satellite data. They are the natural starting point for world models, covering as they do, the whole world.<br>NASA’s Landsat series of satellites began monitoring the whole landmass of the globe from space in 1972. Twice a month it revisited the globe at 30m resolution. Planet’s SuperDove satellites scan nearly every point on Earth’s landmass, at 10x higher resolution, every single day, and have done so since 2017. Together, Landsat and SuperDoves have imaged the Earth ~5,000 times. This is the continuous visual memory of our planet. This is the real-world data equivalent of Wikipedia.<br>And there’s far more data to consider. Real-time camera feeds monitor roads, ports, and coastlines. Weather networks sense atmospheric conditions at millions of points. IoT devices embedded in farms, factories. The physical world has never been more richly instrumented.<br>With Landsat and Planet’s archive of satellite images at its core, this corpus of real-world sensor data is akin to the body of information on the internet. As I said in a 2018 TED Talk , “Google indexed the internet to make it searchable. Planet’s indexing the Earth to make it searchable.” And now, just as AI is unlocking the value of the knowledge on the internet, making it easier to understand, learn, and more versatile and applied, it is now unlocking sensory data about the Earth: accessible and answerable to everyone who needs it.<br>Using space to help life on Earth.<br>We go to orbit not to leave our planet — but to understand it.
I call the models that will learn from this convergence of physical-world data Large Earth Models, or LEMs . Where large language models (LLMs) can tell you what floods look like in general, an LEM would be able to tell you what a specific flood looks like, today, in your town, and how rising water levels compare to every other flood it has ever observed in the region or in similar geographies in other parts of the world. It could cross-reference satellite imagery against rainfall data, road networks, and population density. Where an LLM knows about crop disease from agricultural textbooks, an LEM would have watched it spread live across millions of fields, and would know what actually happened next. While LLMS have enormous potential in the digital realm, 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.<br>The distinction is the difference between awareness and action, because action requires specific, grounded information inside the OODA (Observe, Orient, Decide, Act)...