Google just launched “Earth AI.” Kudos and thank you to Christopher Phillips, James Manyika, and the team. The world is a better place today. They’ve also raised public awareness and set a higher bar for what planetary-scale awareness could and should be, and how AI can help get us there.
Many have asked me for hot takes, since I’m also deep in geoAI. IMO Google also leaked that their geo moat is getting weaker. Their tech paper makes clear this is more about smart orchestration of (open) models than throwing compute at closed data. Earth data — “the other trillion tokens of AI” — is not incremental to text/images but orthogonal.
What the paper actually shows
The three new models are wrappers of existing open models (the RS pillar is SigLIP 2 + MaMMUT with a fine-tune on proprietary data, for an accuracy gain of +2%). The population PDFM embeddings correlate strongly with the mentioned — but not cited — open SatCLIP/GeoCLIP. All Earth AI models, while really good, have open versions not far behind.
Embeddings vs. Agents
The central question is: Is AI for Earth embeddings-based or agents-based? One of the most puzzling choices in Earth AI is abandoning the embedding-based approach that showed so much progress with AlphaEarth (AEF). They wildly downsample their spatiotemporal precision to histogram bins across census-tract levels. Same for PDFM embeddings. No joint latent cross-modal space. “Synergy” here means concatenation with only small or negative gains, with no ablation or correlation studies. For example, the FEMA wildfire risk score drops slightly when both embeddings are used. Across all FEMA tasks, R² difference is just ~0.06. The CDC health showcase’s best results show winter embeddings predicting flu better than July ones (obvious), and the OECD benchmarks show marginal improvement over baseline for GDP of France by extrapolating from countries around it.
These tools aren’t bad — they’re amazing. But they aren’t being used as their own papers instruct. Remarkably, Earth AI involved many parts of Google, but no one from Google DeepMind, the creators of AEF. And AEF has close open peers in Tessera, MOSAIKS, Clay, and others.
Instead, Google “orchestrates” these separate models through a Gemini agent loop, iterating between planning, invoking (inference/fine-tune), and reflecting. Agentic workflows are powerful, but they also add latency, brittleness, emissions, and complexity. Google Earth AI is less multimodal and more a multi-mono-modal MCP loop.
What it means
Google’s Earth AI has a very promising road ahead, and — given the above — so do the rest of us. It validates the importance of the Earth modality of AI. It validates that the competitive edge in geoAI is not so much on closed datasets or deep models, but on putting geo embeddings — open or not — at the center. But for LGND AI’s sake, please keep adding agentic loops based on open models.
The only benchmark that matters is helping on real problems of real people. In that, kudos to Google for Earth AI. It will have tangible benefit.
Originally posted on LinkedIn.