Image: NASA · Artemis II
The problem
Ask Claude or ChatGPT where to protect biodiversity next, or whether a mountain slope will slide tomorrow. Without local data and expert checks, you can’t reliably use the answer.
Models learn from the data that is cheapest and easiest to reach. Earth data comes from hundreds of groups, kept hundreds of different ways.
Our response
Common Crawl is a free, open 12PB archival copy of the web.
Thousands of research studies, then today’s chatbots, were built on it. The Foundation of Modern AI.
PI Pod is building the same for Earth data: a digital library, and an index to find it.
Keep your data your own way. Mirror a copy into the commons.
The library adapts the open framework from Marin, a project led by Percy Liang of Stanford HAI’s Center for Research on Foundation Models.
Research papers citing Common CrawlSource: Common Crawl, Google Scholar citations to 2026
Today
Where we are. The road, and what the avalanche experiment has shown so far.
Where we’re going. Find Earth data and learn from similar places.
What would help us get there. What stands in the way, and how you can help.
Earth data is anything with a place. Today every team collects and aligns it from scratch. We align it once, openly, so places that look alike can learn from each other.
Prove it on avalanche forecasting in Utah.
Layer in more Earth data: what we learn from Planet’s daily imagery, Common Crawl web pages, Earthkind datasets.
Proposed: E3C Data Consortium
Open data and tools to find records and learn from similar places. Not built yet.
UAC experiment · The field test
Photos: Utah Avalanche Center
UAC experiment · The data stack
Every source is resampled onto the same 50 m grid over Little Cottonwood, Big Cottonwood and Mill Creek canyons.
Weather drives a snowpack model in every cell. Terrain, land cover, satellite photos and satellite radar describe the ground it sits on.
UAC experiment · Datasets
Every source we’ve looked at, sorted by how far it got.
UGRCSnowPilot
ClaySLFPANGAEA
Ten winters · 7 UAC forecast regions · ↓ drill in by organization
Datasets · 1 of 6
HRRR · AORC · MADIS · MRMS · NEXRAD
SnowEx · MERRA-2 · OPERA · Prithvi · ICESat-2
NISAR
3DEP · Landsat
SNOTEL
Datasets · 2 of 6
Advisories · avalanche reports · station record
Utah Geospatial Resource Center
snowpilot.org
ESM-SnowMIP · Weissfluhjoch
Atwater · Little Cottonwood
Avalanche Canada · CAIC · NAC · NVE
Datasets · 3 of 6
Sentinel-2 · Sentinel-1 · GLO-30 · WorldCover
ERA5
TESSERA
ALOS-2 · SAOCOM
Datasets · 4 of 6
Planetary Computer
AlphaEarth · ARCO ERA5
Global canopy height
Prithvi
Clay foundation model
Roads, lifts, buildings
Horizon 2 multimodal dataset
Radiant Earth
Datasets · 5 of 6
PlanetScope · Skyscraper · Stories
MesoWest
Open-Meteo · Visual Crossing · meteoblue · Meteomatics
Umbra · ICEYE · Capella
Global news events
Datasets · 6 of 6
Considered, not yet tested. Know one we’re missing? Tell us.
iNaturalist · eBird · BOLD · Xeno-canto · Wildlife Insights · LILA BC · Map of Life
OBIS · Global Fishing Watch
MGnify · NatureMetrics
Maxar Open Data · OpenAerialMap · OpenTopography · SpaceNet · Radiant MLHub
Google Earth Engine · Hugging Face · NASA Earthdata
Our World in Data · WRI · EDGI · Ask Nature
UAC experiment · The results
About as accurate as UAC’s human forecasters, using free public data.
An improvement on UAC’s current snowpack model.
Daily satellite imagery is the piece still to test.
UAC experiment · From stations to every slope
UAC todaySkill = share of avalanche days caught minus share of quiet days flagged (Peirce score), 2025–26 season, 411 avalanches and 488 quiet days.
Next · beyond Utah
We plan to reach new regions through Earth-model biomes: places whose satellite fingerprint looks alike.
Wasatch 2024 · from 10 m GeoTessera embeddings, mapped at 80 m
How better data becomes a better call
Stations, satellites, snowpits. Thousands of readings a day, each one alone.
Every reading aligned to one map: every slope, every 50 m, every day.
What the forecaster now sees between the stations that she could not see before.
The advisory that goes out at 5 am. A person makes it. The model helped her see.
Fewer people in the wrong place on the wrong day.
UAC experiment · Route log
Horizon 1 · Why it matters
If it works, it becomes the proof of concept for H2.
UAC proof of concept as the key that unlocks H2: Multimodal Training Dataset.
Horizon 2 · The Bridge
NextBeyond avalanches: one grid, open from day one.
How PI Pod differs from Amazon.ia: open and free, anywhere on Earth, built from data that already exists.
Web pages and news, about places
Earthkind CollectiveHorizon 3 · Proposed E3C Data Consortium
A shared library would line records up by place and time. Our hypothesis: similar places can reuse what was learned. The index isn’t built yet.
Horizon 3 · Like places, everywhere
01 · Connect the US avalanche centers. Every one of them has the data. None yet has the map.
02 · …then centers worldwide forecast probability
03 · …then other natural hazards
04 · A networked Earth
Cross-coordination
Positioning
Different jobs, and they fit together: Amazon.ia’s results could be one more layer in the PI Pod index.
Amazon.ia launched 23 Sept 2026, led by Planet with more than two dozen partners and Bezos Earth Fund grants. Its public material does not state data-access terms.
Inspiration
In common: learn from the Earth itself, not text alone. Most Earth data already sits in open archives, unused. The goal is stewardship, with people making the call.
Source: “Natural General Intelligence”, naturalgeneralintelligence.ai, Ryan Orbuch (Lowercarbon Capital), 2026.
Why PI exists
Monitoring, feedback, foresight
Gap
Data sovereignty, coalition governance
Long-horizon monitoring, adaptive management
Disaster preparedness, water security
Open question
Why PI exists
Process architecture frame: EarthKind Collective, “WIT v2: Strategic Responses to the Metacrisis”, 2026 · Click a phase, or press 1–4
Current Large Language Models (LLMs) have consumed the written record of human civilization, functioning like scholars locked in a library. They know agronomy, oceanography, or climate science in abstract terms, but remain disconnected from daily physical reality on the ground.
Credit: Will Marshall, “Planetary Intelligence”, Substack 2026
Satellite constellations have provided humanity with extraordinary “disembodied eyes” that scan planetary transformations and emerging threats in real time. However, without an integrated sensory-compute nervous system or responsive institutions, daily observation leaves humanity as a passive witness rather than an active steward.
Credit: Robbie Schingler, “How Satellites & AI Can Protect the Planet”, TED 2026; WEF 2026
Because AI lacks a continuous, real-time baseline model of Earth dynamics, decision-makers alter planetary systems without noticing the cascading feedback loops across climate, oceans, and ecosystems until crises erupt.
Credit: Ryan Orbuch, “Natural General Intelligence”, 2026; Will Marshall, 2026
Planetary Intelligence re-orients AI from text processing toward real-world foundation models trained on continuous satellite imagery and global sensor data.
Credit: Will Marshall, “Planetary Intelligence”, Substack 2026; WEF 2026
Operating on the neuroscience principle of predictive coding, a Planetary Intelligence Model maintains a baseline expectation of normal Earth dynamics. When reality diverges (unexplained smoke, rapid water loss, unannounced troop movements), it generates a “prediction error” that grants humanity proprioception (real-time planetary self-awareness) and executive function (reasoning about downstream risks).
Credit: Marshall, Schingler, & Zolli, World Economic Forum 2026
Connecting neural networks to a high-bandwidth sensory stream of living Earth reality serves as an act of embodiment. Experiencing ecosystems and human livelihoods directly gives AI a reason to value life, serving as an alignment lever and helping humanity navigate the “Great Filter” toward true Planetary Wisdom.
Credit: Will Marshall, “Planetary Intelligence”, Substack 2026
Instead of downlinking hundreds of petabytes of raw satellite imagery to Earth, AI processors (NVIDIA chips and Google TPUs) are being launched directly into orbit to process imagery at the sensor, taking advantage of continuous 24/7/365 polar solar power.
Credit: Will Marshall, Substack 2026; WEF 2026; Planet Labs & Google, 2026
Uniting 50+ years of satellite archives (NASA Landsat, Planet SuperDoves) with real-time in-situ sensors (IoT, weather networks, bioacoustics, and eDNA) into curated, durable environmental training sets.
Credit: EarthKind Collective, “Planetary Intelligence Initiative”, 2026; Ryan Orbuch, NGI 2026
Establishing open-data infrastructure and feedback loops between data producers and model developers (e.g., Common Crawl, OpenAthena) ensures training sets are openly governed, high-quality, and structurally ready for pre-training.
Credit: EarthKind Collective, “Planetary Intelligence Initiative”, 2026
The ultimate purpose of Planetary Intelligence is not static analysis or academic reports, but closing the gap between awareness and action inside the human OODA loop (Observe, Orient, Decide, Act).
Credit: Will Marshall, Substack 2026; Ryan Orbuch, NGI 2026
Information alone does not save ecosystems or prevent conflict—only coordination does. While satellites provide the eyes and space compute provides the brain, human governance, trust, and alliances must form the body brave enough to move when the system signals an emergency.
Credit: Robbie Schingler, TED 2026
Deploying real-world models to drive concrete interventions across key domains—such as real-time wildfire tracking, dynamic insurance underwriting, detecting unpermitted construction, enforcing maritime security against dark ships, and managing disaster response before cascades occur.
Credit: Will Marshall, Substack 2026; Robbie Schingler, TED 2026; EarthKind Collective, WIT v2 2026
What would help us get there
Three things slow us down. Clearing any one of them speeds up the experiment.
Daily imagery. We need Planet data to answer the satellite question.
Capacity. A small team; a few advisors would go a long way.
Permission. Agreeing with data partners how results can be shared.
Where you can help
Two or three people to help us think through the next step.
Funders open to technology and the environment, and a way to Planet’s daily imagery.
Know a dataset with a place and a time? Add it to our list.
If your work produces data with a place and a time, or you know who holds it, find us at dinner.