Ai2 released OlmoEarth v1.3 and precomputed 10-meter satellite embeddings for the whole planet
AspireAI community · Allen Institute for AI (Ai2)
*One satellite model, and the planet already indexed.*
OlmoEarth v1.3 reads a year of Sentinel-1 radar plus Sentinel-2 and Landsat imagery and turns every 10-meter pixel of Earth into a 128-dimensional vector (the first 64 dimensions work standalone). Ai2 published the model weights, the pretraining data and the pretraining code, and also computed the embeddings for the whole planet - roughly 150 TB per year, global for 2023-2025 with select-country coverage back to 2017 - so a crop, land-cover or change map is now a small classifier trained on top instead of a fine-tuning run.
In Ai2's evaluation against AlphaEarth Foundations and Tessera v2 across 27 evaluations on six continents, the 128-dimensional embeddings scored highest on 20 of 27 and the 64-dimensional version still led 18 of 27. They also introduced PLANTEUR, a tropical crop-type benchmark, and found the embeddings generalise best to places a classifier has never seen. It was their cheapest model yet: 53% of v1.2's compute, and the final pretraining run cost about $2,720 in rented H100 node time versus $15,060 for v1.
Caveat: those cost figures cover the final pretraining run only - not experiments, data preparation, storage or engineering time - and the benchmark comparison is Ai2's own.
https://huggingface.co/blog/allenai/olmoearth-v1-3
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.